Text abstract illusion suppression method based on generation optimization and post-processing collaboration

Abstract

The invention belongs to the technical field of natural language processing and text generation, and relates to a text abstract illusion suppression method based on generation optimization and post-processing collaboration. The method comprises the steps of inputting a source text, generating an initial abstract by adopting a model trained by a decoding strategy based on domain condition point mutual information, and inhibiting generation of irrelevant words introduced by domain deviation by comparing conditional probabilities and marginal probabilities of generated words in a source text domain so as to improve abstract loyalty; and the post-processing framework constructs non-fact abstract training data containing common errors by utilizing a filling model, drives a correction model to perform multi-dimensional detection and correction on the initial abstract in combination with a self-checking mechanism, covers dimensions such as entities, quantity, negative and semantic consistency, and finally outputs a high-fidelity abstract. According to the method, collaborative optimization between deviation suppression in the generation stage and global correction in the post-processing stage is realized, the illusion rate is reduced, the abstract fluency is kept, and the abstract reliability is improved.

Description

Technical Field

[0001] This invention belongs to the field of natural language processing and text generation technology, and relates to a method for suppressing text summarization illusions based on the synergy of generation optimization and post-processing. Background Technology

[0002] With the rapid development of the internet, information has become the most dynamic resource, and its output is increasing exponentially. How to extract useful data from massive amounts of information for users has become an urgent problem to solve, and text summarization technology has filled this gap. Text summarization technology is an information compression technique that utilizes computers and corresponding rules, algorithms, and models. Its aim is to quickly extract a concise, coherent, and comprehensive short text that reflects the core content of a document, while being shorter than the original. Early text summarization relied mainly on manual operation, but with the increase in unstructured text in recent years, manually processing such unstructured text has become extremely tedious. Text summarization technology can effectively compensate for the shortcomings of manual processing; therefore, text summarization technology has become a hot topic in natural language processing research.

[0003] Early explorations of Automatic Language Summarizing (ATS) systems can be traced back to the 1950s. Subsequently, in the 1990s and early 2000s, with the advancement of statistical machine learning, systems based on unsupervised features began to emerge. By the 2010s, the focus of summarizing research shifted to training deep learning models under supervised frameworks using large amounts of training data. More recently, the emergence of self-supervised pre-trained language models (PLMs) like BERT and T5 has significantly improved summarizing performance through a "pre-train, then fine-tune" process. This progress has finally ushered in a new era dominated by large language models (LLMs). A review of the development history of summarizing methods can be roughly divided into four stages: the statistical method-based stage, the deep learning stage, the pre-trained language model fine-tuning stage, and the current large language model stage.

[0004] Text summarization technology has evolved from simple extraction-based summarization using statistical methods to generative models in the deep learning stage, then to fine-tuning of pre-trained language models (such as BERT and GPT), and finally to the current stage of large-scale language models (LLM). With this technological evolution, the quality of generated summaries has significantly improved, especially in areas such as language fluency, semantic understanding, and content compression. However, despite these advancements, current automatic summarization technologies still face some problems and challenges.

[0005] The illusion problem in text summarization tasks mainly manifests as the model-generated summary containing facts or information not found in the original text, or incorrectly summarizing and generalizing the original content. This false information may appear logical, fluent, and even related to parts of the original text, but in reality, it fails to accurately reflect the true meaning or content of the original. Specifically, the complete fictional illusion refers to the model-generated summary containing false information unrelated to the original text, while the semantic deviation illusion refers to incorrect summarization of the original content. For example, if the original text discusses climate change, the summary might add financial forecasts out of thin air; or incorrectly summarize "global warming is intensifying" as "warming has been curbed." This phenomenon stems from the fact that language models rely on word probabilities rather than semantic understanding in their generation mechanism, leading to seemingly reasonable but factually inaccurate content. Furthermore, the model is prone to ignoring key details or introducing erroneous information during information compression, and the combined effects of training data quality defects (such as noise and domain mismatch) and insufficient logical reasoning ability further exacerbate the illusion risk.

[0006] Such problems have a particularly significant impact on high-precision fields such as medicine and law: false information may lead to misdiagnosis or legal misjudgment, resulting in serious consequences; frequent hallucinations can undermine users' trust in automated systems; and ethical and liability risks (such as misleading news reports) are more likely to trigger legal disputes. Therefore, it is necessary to systematically reduce the hallucination rate and ensure the accuracy and practical application value of summaries by optimizing generation strategies to suppress domain bias, enhancing post-processing correction capabilities, and combining self-checking mechanisms to improve semantic consistency.

[0007] Methods for detecting text hallucinations lack universality. Existing methods for detecting hallucinations in text summarization mainly fall into two categories: task-specific methods and general methods, each with its own characteristics and applicable boundaries. Task-specific methods focus on single-domain (e.g., medical, legal) design, using domain-annotated data to enhance the model's ability to identify specific hallucination patterns (e.g., medical entity errors, legal clause misinterpretations), demonstrating high accuracy in vertical scenarios. For example, medical domain models can effectively detect fictitious content in diagnostic summaries through terminology matching and symptom logic verification. However, these methods heavily rely on domain knowledge bases and high-quality annotated data, facing problems such as poor cross-domain generalization and high data acquisition costs, and are difficult to handle novel hallucination patterns in complex tasks (e.g., cross-modal summarization or multi-document integration).

[0008] General-purpose methods enhance universality by optimizing the generation mechanism (e.g., introducing semantic alignment loss, enhancing contextual reasoning) or post-processing corrections (e.g., rule-based validation, multi-model collaborative validation), and are not limited to specific domains. For example, contrastive learning-based strategies can suppress fictitious and deviant phenomena in multi-domain summarization by measuring the semantic consistency between the generated content and the original text. The advantage of such methods is that they do not require domain-specific prior knowledge and can achieve cross-task adaptation with only general corpora, making them more flexible, especially in scenarios with scarce data or diverse tasks. However, their detection accuracy may be lower than domain-customized models, and their ability to capture complex logical errors (such as distortions in medical causality) is limited. In practical applications, these two types of methods often need to be combined, achieving a balance between hallucination detection accuracy and generality through the synergy of domain-adaptive training and a global correction framework. Summary of the Invention

[0009] To address the lack of universality in current methods for detecting and suppressing text hallucinations in text summarization, this paper proposes a text summarization hallucination suppression method based on the synergy of generation optimization and post-processing. This method aims to simultaneously solve the problems of hallucination detection and reduction. Specifically, it utilizes a domain-conditional pointwise mutual information decoding strategy and a post-processing framework based on irrelevant models to design a solution applicable across domains and tasks.

[0010] A text summarization illusion suppression method based on the synergy of generation optimization and post-processing is characterized by the following steps:

[0011] S1. Obtain the original dataset, including the source text and its reference summary; divide the dataset into training and test sets; preprocess the data in the training and test sets.

[0012] S2. Input the preprocessed training set data into the BART-large model. During the decoding process, a domain-conditional strategy is used to optimize the generation process, resulting in an optimized summary generation model. ;

[0013] S3. Input the data from the test set into the trained summary generation model. In the process, an initial summary is obtained;

[0014] S4. Perform error-prone term identification and cloze test sample construction on the preprocessed training set data to obtain the cloze test question set data.

[0015] S5. Input the cloze test question set into the BART-large model. During the training phase, the model learns to predict the masked words given the context, ultimately obtaining the trained cloze test model. ;

[0016] S6. Input the cloze test question set data into the cloze model. Negative sample non-fact summary data is generated in the middle;

[0017] S7. Input the preprocessed training data of positive samples and the non-fact summary data of negative samples into the Flan-T5 model. Combined with instructional prompts, the model learns how to distinguish between normal and false information and corrects the false parts through training, finally obtaining a trained illusion correction model. ;

[0018] S8. Input the initial summary and its source text into the hallucination correction model. Combined with instructional prompts, output the hallucination judgment result and the corrected summary;

[0019] The beneficial effects of this invention are:

[0020] (1) The Domain Conditional Strategy (PMIDC) of the present invention can significantly improve the consistency between the generated content and the source text by subjecting the probability of generating tokens to domain conditions, thereby effectively reducing the occurrence of illusion phenomena. In particular, when dealing with high-entropy generated content, it can more accurately control the uncertainty in the generation process.

[0021] (2) The post-processing framework of the present invention, especially the hallucination correction model, improves the authenticity and credibility of the generated summary by automatically identifying and correcting hallucinations in the summary, and provides a more accurate solution for text summarization in practical applications. Attached Figure Description

[0022] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0023] Figure 1 This is an overall flowchart of an embodiment of the present invention;

[0024] Figure 2 This is a schematic diagram illustrating the construction of a cloze test according to an embodiment of the present invention;

[0025] Figure 3 This is a schematic diagram of the filling model structure according to an embodiment of the present invention;

[0026] Figure 4 This is a schematic diagram of the illusion correction model structure according to an embodiment of the present invention; Detailed Implementation

[0027] 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.

[0028] Figure 1 The diagram shown is an overall flowchart of an embodiment of the present invention. Figure 1 As shown, the implementation process of the present invention includes a generation stage and a post-processing stage: in the generation stage, the source text is input, and an initial summary is generated by a model trained by a decoding strategy based on domain condition point mutual information; in the post-processing stage, the initial summary is detected and corrected by using a filling model and an illusion correction model, and finally a high-fidelity summary is output.

[0029] This embodiment provides a text summarization illusion suppression method based on the synergy of generation optimization and post-processing, including but not limited to the following implementation process:

[0030] During the generation phase, the input source text is used. The model, trained based on the domain conditional mutual information decoding strategy, is optimized by the domain conditional mutual information scoring function to suppress the illusion caused by domain bias.

[0031] The CNN / DailyMail text summarization dataset is used as the base dataset. This dataset consists of news articles from CNN News and the Daily Mail, each containing the original article and manually written reference summaries (highlights), totaling 60,000 news-summary pairs. The dataset is divided into three parts: a training set containing 287,227 news-summary pairs; a test set containing 13,368 news-summary pairs; and a validation set containing 11,490 news-summary pairs.

[0032] The BART-large model is trained using the training set from the base dataset. First, the training set is preprocessed, including cleaning to remove irrelevant characters, redundant spaces, unusual symbols, and non-textual content, resulting in preprocessed source text. The summary generation process uses the input source text as input. ,in For the source text set; given the source text Under the given conditions, the source text is compressed and a corresponding output summary sequence is generated. ,in The target sequence set; the output summary sequence From the vocabulary list The sequence is composed of tokens, starting with the first token. Start, to terminate token End, and the sequence length is PMIDC scoring is used to calculate the score for each token generated. The scoring function is calculated as follows:

[0033]

[0034] in, Indicates that in the given source text and partially generated sequences Generate token under the following conditions The conditional probability; This indicates that the source text is not considered. In the case of only the partial sequence Generate token The conditional probability; These are hyperparameters used to control the strength of the penalty term; It is a binary indicator used to indicate the current generation step. Whether to enable penalty items, and if the following conditions are met:

[0035]

[0036] in, Indicates an indicator function; The model represents the steps The predicted distribution entropy; A preset threshold is used; when the entropy is greater than the threshold... season Otherwise ;

[0037] In the post-processing stage, hallucination detection and correction are performed. The initial summary is input into the post-processing framework, which includes cloze test construction, a completion model, and a hallucination correction model. The specific steps are as follows:

[0038] Construct a cloze test. Figure 2 This is a schematic diagram illustrating the construction of a cloze test according to an embodiment of the present invention. Input training source text. and its corresponding reference abstract Tools such as OpenIE are used to identify all error-prone terms in the text and create a set. OpenIE is primarily used to extract structured information such as subject-verb-object triples from unstructured text. The relational information extracted by OpenIE is then analyzed to identify terms that may lead to misunderstandings. These terms may include polysemous words, technical terms, or any words that do not conform to the expected format. Subsequently, each identified error-prone term is... Replace with one The markers are used to create a cloze test problem; this process is formalized as a function. Its output is a set of cloze test questions, denoted as . :

[0039]

[0040]

[0041] Where the function Indicates: Text Terms in Replace with "[MASK]" and return the modified text. For example, in the summary “The company reported a profit of $5 million”, “$5 million” is replaced with “The company reported a profit of [MASK]”.

[0042] Fill in the model training and inference. Figure 3 This is a schematic diagram of the infill model structure according to an embodiment of the present invention. Infill Model It is trained from the pre-trained masked language model BART-large. During the training phase, the input is a sentence with [MASK]. and training source text The target output is the prediction of each The correct term is masked at the specified position. The loss function is cross-entropy loss, and the optimizer uses AdamW. During the inference phase, beam search is used to generate multiple candidate terms (in sequence generation tasks, the model typically needs to select a word from the vocabulary at each step until a complete sequence is generated. Beam search retains only the k most likely candidate sequences at each time step and continues to expand based on these candidate sequences until an end-of-sequence symbol is generated or the maximum length is reached). Candidate terms with lower prediction probabilities are usually similar to the original terms semantically or in part-of-speech, but lack factual accuracy. Therefore, candidate terms with lower prediction probabilities are selected as non-factual replacements for the masked terms, used to construct a non-factual summary, denoted as . For example, if the original term is "$5 million", the model might generate "$10 million" as an illusion sample.

[0043] Illusion correction model training and inference. Figure 4 This is a schematic diagram of the illusion correction model structure according to an embodiment of the present invention. (Filling model) It is trained from the Flan-T5 model. During the training phase, a positive and negative sample pair is constructed to correct the hallucination. Supervised training was conducted with 40% positive samples, using a reference summary (without hallucinations) as input, and the target output being... Tags and corresponding summary text; negative samples account for 60%, using the illusion summary generated by the filler model as input, and the target output is... Tagged and corrected summary text.

[0044] The input format for the hallucination correction model is the instruction prompt "Check for any hallucination in the summary, and if found, correct them.", the initial summary, and its source text. The target output format is " "+ Corrected Summary". The initial summary is the text to be detected and corrected. This indicates that there was no hallucination. This indicates that a hallucination has been detected.

[0045] For example, when the input is "Source: [Source Text]. Summary: [Initial Summary]", the hallucination correction model can output " The company reported a profit of $5 million (corrected from $10 million).”

[0046] When describing elements of various embodiments of this application, the articles “a,” “an,” “this,” and “described” are all intended to indicate that there are one or more elements. The words “comprising,” “including,” and “having” are inclusive and mean that there may be other elements in addition to those listed.

[0047] It should be noted that those skilled in the art will understand that all or part of the processes in the above method embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0048] The above description is merely a specific embodiment of this application. It should be noted that those skilled in the art will understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A hallucination suppression method for text summarization based on generation optimization and post-processing synergy, characterized in that, Includes the following steps: S1. Obtain the original dataset, including the source text and its reference summary; divide the dataset into training and test sets; preprocess the data in the training and test sets. S2, input the preprocessed training set data into the BART-large model, and adopt the domain condition strategy in the decoding process to optimize the generation process to obtain an optimized summary generation model ; S3. Input the data from the test set into the trained summary generation model. In the process, an initial summary is obtained; S4. Perform error-prone term identification and cloze test sample construction on the preprocessed training set data to obtain the cloze test question set data. S5. Input the cloze test question set into the BART-large model. During the training phase, the model learns to predict the masked words given the context, ultimately obtaining the trained cloze test model. ; S6. Input the cloze test question set data into the cloze model. Generate negative sample non-fact summary data; S7. Input the preprocessed training data of positive samples and the non-fact summary data of negative samples into the Flan-T5 model. Combined with instructional prompts, the model learns how to distinguish between normal and false information and corrects the false parts through training, finally obtaining a trained illusion correction model. ; S8. Input the initial summary and its source text into the hallucination correction model. Combined with instructional prompts, it outputs the hallucination judgment result and a corrected summary.

2. The text summarization hallucination suppression method based on the synergy of generation optimization and post-processing as described in claim 1, characterized in that, The preprocessing of the original data dataset includes: The source text and reference abstract are cleaned to remove irrelevant characters, extra spaces, abnormal symbols, and non-text content; the text is standardized, including unifying capitalization, correcting spelling errors, and standardizing abbreviations and number formats; the text is segmented and stop words are removed.

3. The text summarization hallucination suppression method based on the synergy of generation optimization and post-processing as described in claim 1, characterized in that, The preprocessed training data is input into the BART-large model. During the decoding process, a domain-conditional strategy is used to optimize the generation process, resulting in an optimized summary generation model. ,include: During the BART-large model decoding process, the input data is the source text from the preprocessed training set data. ,in For the source text set; given the source text Under the given conditions, the source text is compressed and a corresponding output summary sequence is generated. ,in The target sequence set; the output summary sequence From the vocabulary list The sequence is composed of tokens, starting with the first token. Start, to terminate token End, and the sequence length is ; Based on the source text The optimal output summary sequence is determined by maximizing the evaluation function. The optimal output summary sequence The calculation expression is as follows: ； in, Indicates that in the given source text Under the condition of [condition], for candidate output summary sequences The rating value; This refers to "the value of the independent variable that maximizes the objective function", that is, within the candidate set. Selected The candidate output summary sequence that achieves the maximum score is taken as the optimal output summary sequence. ; To search for and obtain high-scoring output summary sequences in the candidate set, beam search is used to approximate a traversal of possible sequences. At each time step, beam search retains only the k most likely candidate sequences and continues to expand based on these candidate sequences until an end-of-sequence symbol is generated or the maximum length is reached. Furthermore, to mitigate beam search's bias towards short sequences, length normalization can be introduced into the sequence scoring, and the evaluation function... The summation of scores at each time step is defined by the following expression: ； in, This represents the sequence score after length normalization; Represents the source text; Indicates the candidate output sequence; Indicates the generation time step; This indicates the final number of steps in generating the output sequence; Indicates the first A token generated at each time step; Indicates the first The partial sequence already generated before the step; This represents the length penalty coefficient, used to adjust the strength of length normalization; For length normalization; Furthermore, a domain-conditional strategy is employed for optimization during the BART-large model decoding process, and then the model's conditional log-likelihood loss function is minimized. For the parameters of the BART-large model The model is updated to learn and generate sequences that more closely resemble the reference summary, ultimately resulting in an optimized summary generation model. The loss function The calculation expression is as follows: ； in, The parameters representing the BART-large model The training loss function; This represents the set of all trainable parameters of the BART-large model; This refers to the BART-large model; This represents a pair of samples in the training dataset. For source text, A reference summary corresponding to the source text; Indicates in the given source text Generate the entire target summary sequence under the conditions The conditional probability; This indicates that the summation is performed on all pairs of samples in the training dataset.

4. The text summarization hallucination suppression method based on the synergy of generation optimization and post-processing as described in claim 3, characterized in that, The BART-large model employs a domain-conditional strategy during decoding, including: The Domain-Conditioned Strategy (PMIDC) reduces the model's over-reliance on domain-related terms when uncertainty is high by performing domain-conditional modeling on the probability distribution of generated tokens, thereby reducing the correlation with the source text. Inconsistent generated content; specifically, PMIDC incorporates domain information into its single-step scoring. Furthermore, both the conditional probability term and the penalty term based on the input are domain-conditionalized, and the calculation expression is as follows: ； in, This represents the domain information corresponding to the source text; The first sentence of the source text is chosen because it usually contains the key information of the text. Indicates that in the given source text The aforementioned domain information and the generated partial sequences Generate token under the following conditions The conditional probability; Indicates information in the given domain and the aforementioned partial sequence Generate token under the following conditions The conditional probability; These are hyperparameters used to control the strength of the penalty term; It is a binary indicator used to indicate the steps. Whether to enable the penalty item is calculated using the following expression: ； in, Entropy is used to characterize the process in the step. Uncertainty; It represents the variance of the domain conditional probability distribution, used to characterize the stability of predictions on this token; These are hyperparameters used to balance the entropy and variance terms; This is a preset threshold; when the uncertainty metric within the parentheses exceeds the threshold... season Otherwise ; Furthermore, regarding the variance term... We propose an achievable estimation method that uses a small model, GPT-2, for K randomized sampling during the training phase. Each sampling obtains different forward propagation results by activating the Dropout layer of the model, thereby obtaining K different output probability distributions. ； in, Indicates the number of randomized samples; Indicates the first The GPT-2 model on tokens under sub-randomized sampling The predicted probability; express token under the second sampling The average predicted probability; Represents token exist The variance of the predicted probability under each sampling; Furthermore, the variance estimates can be aggregated into a scalar uncertainty index for gating decisions, such as for top- The sum of the variances of the tokens is calculated using the following expression: ； in, Indicates by The first one with the highest probability after sorting A set of tokens; This represents the aggregated variance scalar, used together with the entropy term to form the gating condition.

5. The text summarization hallucination suppression method based on the synergy of generation optimization and post-processing as described in claim 1, characterized in that, Input the data from the test set into the trained summary generation model. In the process, an initial summary is obtained, including: The source text from the preprocessed test set data is sequentially input into the optimized summary generation model. During the decoding and generation process, the domain-conditional strategy PMIDC is used to optimize the token generation at each step, in order to generate a summary sequence with higher consistency with the source text; finally, the summary generation model... Generate an initial summary.

6. The text summarization hallucination suppression method based on the synergy of generation optimization and post-processing as described in claim 1, characterized in that, The preprocessed training set data is used to identify common error terms and construct cloze test samples, resulting in a set of cloze test questions, including: Input training dataset source text and its corresponding reference abstract Tools such as OpenIE are used to identify all error-prone terms in the text and create a set. OpenIE is primarily used to extract structured information such as subject-verb-object triples from unstructured text. The relational information extracted by OpenIE is used to analyze terms that may lead to misunderstandings; these terms may include polysemous words, technical terms, or any words that do not conform to the expected format. Subsequently, each identified error-prone term is replaced with a... The markers are used to create a cloze test problem; this process is formalized as a function. Its output is a set of cloze test questions, denoted as . : ； ； Where the function Indicates: Text Terms in Replace with "[MASK]" and return the modified text. .

7. The text summarization hallucination suppression method based on the synergy of generation optimization and post-processing as described in claim 1, characterized in that, The cloze test question set is input into the BART-large model. During the training phase, the model learns to predict the masked words given the context, ultimately resulting in a trained cloze test model. ,include: The input to the BART-large model consists of... Marked sentences Its training source text The results are obtained by concatenating the data, and the target output is the prediction of each... The correct term is occluded at the specified location; through training, the model gradually learns to recover the occluded term based on the context, ultimately resulting in a well-trained filling model. .

8. The text summarization hallucination suppression method based on the synergy of generation optimization and post-processing as described in claim 1, characterized in that, Input the set of cloze test questions into the cloze model. The generation of negative sample non-fact summary data includes: During the inference phase, the aim is to generate a non-factual summary that has slight deviations from the reference summary. During inference, the trained filler model... A beam search method is used to generate multiple candidate terms. Candidate terms with lower prediction probabilities are usually similar to the original terms semantically or in part of speech, but lack factual accuracy. Therefore, candidate terms with lower prediction probabilities are selected as non-factual replacements for the masked terms; the non-factual summary is denoted as... Its generation method is: replacing a candidate token with a different one. In mark; In sequence generation tasks, the autoregressive model BART-large typically needs to select a word from the vocabulary at each step until a complete sequence is generated; while bundle search retains only the k most likely candidate sequences at each time step and continues to expand based on these candidate sequences until an end symbol is generated or the maximum length is reached.

9. The text summarization hallucination suppression method based on the synergy of generation optimization and post-processing as described in claim 1, characterized in that, The preprocessed training data of positive samples and the non-fact summary data of negative samples are input into the Flan-T5 model. Combined with instructional prompts, the model learns how to distinguish between normal and false information and corrects the false parts through training, ultimately resulting in a well-trained illusion correction model. ,include: An imperative prompt, "Check for any hallucination in the summary, and if found, correct them.", is constructed to guide the Flan-T5 model to perform hallucination detection before generating the corrected summary, and a self-checking flag is introduced. This indicates that hallucinations are present. This indicates that no hallucination has been detected; the input to the Flan-T5 model consists of the instruction prompt "Check for any hallucination in the summary, and if found, correct them." and the source text. Together with positive and negative sample summaries, the target output of the Flan-T5 model includes the self-testing markers. or and the corresponding revised summary; During the training phase, the illusion correction model is trained under supervision using both positive and negative samples, with positive samples accounting for 40% and negative samples accounting for 60%. The positive samples are derived from reference summaries, and the negative samples are filled by the model. The generated non-fact summary; finally, a trained illusion correction model is obtained. .

10. A text summarization hallucination suppression method based on generation optimization and post-processing synergy as described in claim 1, characterized in that, The initial summary and its source text are input into the hallucination correction model. Based on the imperative prompts, output the hallucination judgment result and the corrected summary, including: A data set is constructed that includes the imperative prompt "Check for any hallucination in the summary, and if found, correct them.", an initial summary, and its source text, serving as a hallucination correction model. The model takes the input from the source text as input and identifies and locates non-factual statements, unverifiable or contradictory illusory content in the initial summary by comparing the content of the source text, and generates self-verification tags. or ;in This indicates that hallucinations are present. This indicates that no hallucination was detected; for those identified as The model corrects, rewrites, or deletes the initial summary based on the contextual information of the source text, generating a semantically consistent and factually accurate corrected summary text; finally, the model outputs the hallucination judgment result and the corrected complete summary text.

Images