Multimodal affective hallucination assessment and mitigation methods, systems, devices, and storage media

Abstract

This invention discloses a multimodal emotional hallucination assessment and mitigation method, system, device, and storage medium, which are corresponding solutions. In these solutions: the emotional understanding process is structured into a perception layer and a reasoning layer, achieving a fine-grained characterization of complex emotional reasoning processes. Furthermore, an atomization mechanism is introduced to decompose the generated complex text into multiple independent minimum semantic units, allowing each emotional judgment or reasoning step to be evaluated individually. This fine-grained decomposition method solves the problem of traditional overall assessment's difficulty in locating the source of errors, providing a foundation for subsequent accurate correction. A task routing strategy matches the most suitable assessment model to different types of minimum semantic units, significantly improving the accuracy of hallucination recognition. In addition, a closed-loop mechanism of assessment-feedback-correction-memory is introduced during the generation process, thus avoiding the error accumulation problem in traditional post-processing methods.

Description

Technical Field

[0001] This invention relates to the field of multimodal large model and emotion computing technology, and in particular to a method, system, device and storage medium for assessing and alleviating multimodal emotional illusions. Background Technology

[0002] With the development of multimodal large language models, the integration of multiple information sources such as text, speech, images, and video for unified understanding and generation has become an important research direction in the field of artificial intelligence. In the field of affective computing, traditional methods mainly rely on classification models to identify predefined emotion labels, such as modeling facial expression recognition, speech emotion recognition, or text emotion analysis separately, and then obtaining the final result through a simple fusion strategy. The advantage of this type of method is that it has a clear structure and strong interpretability, but its limitations are also very obvious. That is, it relies on a fixed label system, has difficulty handling open domains, multi-granularity, and complex emotion expressions, and cannot meet the ever-changing emotion understanding needs in real-world scenarios.

[0003] With the development of generative artificial intelligence, multimodal large language models are increasingly being used for open-ended sentiment understanding tasks. These models no longer output fixed labels but instead generate sentiment interpretations in natural language form. This approach is closer to human expression and possesses stronger generalization capabilities. However, these models commonly suffer from the "illusion" problem, where the model generates sentiment judgments that are inconsistent with or lack supporting evidence compared to the actual input. For example, the model might infer a person's emotions out of thin air, exaggerate intonation, or engage in complex psychological reasoning without sufficient evidence. This problem is particularly prominent in sentiment understanding tasks because sentiment reasoning itself is subjective and multi-layered; once an error occurs, it can easily be amplified in subsequent reasoning, creating a "snowball effect" of errors.

[0004] Existing solutions to the hallucination problem mainly fall into two categories: post-processing verification methods, which verify and correct errors after generation using rules or additional models; and generation-stage optimization methods, such as adjusting decoding strategies or introducing contrastive learning mechanisms to reduce error probability. However, these methods have significant shortcomings. Post-processing methods cannot prevent the propagation of errors during the generation process and can only perform post-processing repairs; while generation optimization methods typically lack fine-grained control and are difficult to precisely intervene at specific semantic units. Furthermore, existing methods generally lack modeling of the "emotional reasoning structure," failing to distinguish between errors at the perception and reasoning levels, resulting in limited optimization effects. Therefore, how to achieve real-time identification and dynamic control of emotional hallucinations during the generation process has become a critical technical problem that urgently needs to be solved.

[0005] In view of this, the present invention is hereby proposed. Summary of the Invention

[0006] The purpose of this invention is to provide a method, system, device, and storage medium for assessing and mitigating multimodal emotional hallucinations. By constructing an assessment-driven closed-loop mechanism, emotional hallucinations are identified and suppressed in real time during the generation process, thereby avoiding the problem of error accumulation in traditional post-processing methods and improving prediction performance.

[0007] The objective of this invention is achieved through the following technical solution: A multimodal method for assessing and alleviating affective hallucinations includes: Inference is performed on multimodal data based on a multimodal large language model to generate original sentiment interpretation text containing multiple sentences; Each sentence is decomposed into its smallest semantic unit, and each smallest semantic unit is classified into a predetermined sentiment analysis dimension; the sentiment analysis dimension includes a perception layer and a reasoning layer dimension. Based on the sentiment analysis dimension of each minimum semantic unit, an evaluation model is matched according to the task routing strategy. The evaluation model determines whether the corresponding minimum semantic unit is a hallucination. When it is determined to be a hallucination, the position of the corresponding minimum semantic unit and the position of the corresponding hallucination word are stored in the hallucination archive. Based on whether all minimum semantic units are hallucinations, the hallucination rate of each sentence is calculated. For each sentence, if the hallucination rate is not higher than the threshold, it is directly taken as the optimal result; if the hallucination rate is higher than the threshold, K rounds of iterative correction are performed, and the hallucination words are suppressed by referring to the information in the hallucination archive during the iterative correction process. For sentences undergoing K rounds of iteration for correction, the sentence with the lowest illusion rate is selected as the optimal result from the K rounds of iterations. The corrected sentiment interpretation text is then constructed based on the optimal results of all sentences and stored in the anchor archive. The final sentiment interpretation text is obtained by re-inferring based on the multimodal large language model combined with the input multimodal data and the anchor archive.

[0008] A multimodal affective hallucination assessment and mitigation system, used to implement the aforementioned method, includes: The input processing module is used to perform reasoning on multimodal data based on a multimodal large language model and generate original sentiment interpretation text containing multiple sentences. The multidimensional evaluation module is used to decompose each sentence into its smallest semantic unit and classify each smallest semantic unit into a predetermined sentiment analysis dimension; the sentiment analysis dimension includes a perception layer and an inference layer dimension. The intelligent agent routing module is used to match the corresponding evaluation model based on the sentiment analysis dimension of each minimum semantic unit using a task-based routing strategy. The evaluation model determines whether the corresponding minimum semantic unit is a hallucination. When it is determined to be a hallucination, the position of the corresponding minimum semantic unit and the position of the corresponding hallucination word are stored in the hallucination archive. Based on whether all minimum semantic units are hallucinations, the hallucination rate of each sentence is calculated. The memory reasoning module is used to directly take the optimal result for each sentence if the hallucination rate is not higher than the threshold; if the hallucination rate is higher than the threshold, it performs K rounds of iterative correction, and during the iterative correction process, it refers to the information in the hallucination archive to suppress hallucination words. The output module is used to select the sentence with the lowest illusion rate from the K rounds of iterations as the optimal result for sentences that have undergone K rounds of iterations. The optimal results of all sentences are used to form the revised sentiment interpretation text and store it in the anchor archive. Based on the multimodal large language model, the input multimodal data and the anchor archive are combined to perform inference again to obtain the final sentiment interpretation text.

[0009] A processing device includes: one or more processors; and a memory for storing one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned method.

[0010] A readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method.

[0011] As can be seen from the technical solution provided by the present invention, the emotion understanding process is structured into a perception layer and a reasoning layer, which achieves a fine characterization of the complex emotion reasoning process. Furthermore, an atomization mechanism is introduced in this process to decompose the generated complex text into multiple independent minimum semantic units, so that each emotion judgment or reasoning step can be evaluated individually. This fine-grained decomposition method solves the problem that traditional overall evaluation is difficult to locate the source of error, providing a foundation for subsequent accurate correction. By matching the most suitable evaluation model to different types of minimum semantic units through a task routing strategy, the accuracy of hallucination recognition is significantly improved. In addition, a closed-loop mechanism of "evaluation-feedback-correction-memory" is introduced in the generation process, thereby avoiding the problem of error accumulation in traditional post-processing methods. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 A flowchart of a multimodal emotional hallucination assessment and mitigation method provided in an embodiment of the present invention.

[0014] Figure 2This is a schematic diagram of the overall framework of a multimodal emotional hallucination assessment and mitigation method provided in an embodiment of the present invention.

[0015] Figure 3 This is a schematic diagram of a multimodal emotional hallucination assessment and mitigation system provided in an embodiment of the present invention.

[0016] Figure 4 This is a schematic diagram of a processing device provided in an embodiment of the present invention. Detailed Implementation

[0017] 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 protection scope of the present invention.

[0018] First, the following explanations are provided for the terms that may be used in this article: The terms "comprising," "including," "containing," "having," or other similar semantic descriptions should be interpreted as non-exclusive inclusion. For example, including a technical feature element (such as raw material, component, ingredient, carrier, dosage form, material, size, part, component, mechanism, device, step, process, method, reaction conditions, processing conditions, parameter, algorithm, signal, data, product or article of manufacture, etc.) should be interpreted as including not only the expressly listed technical feature element, but also other technical feature elements that are not expressly listed and are well-known in the art.

[0019] The following provides a detailed description of a multimodal emotional hallucination assessment and mitigation method, system, device, and storage medium provided by this invention. Contents not described in detail in the embodiments of this invention are prior art known to those skilled in the art. Where specific conditions are not specified in the embodiments of this invention, they are performed according to conventional conditions in the art or conditions recommended by the manufacturer. Where the manufacturers of the instruments used in the embodiments of this invention are not specified, they are all conventional products that can be purchased commercially.

[0020] Example 1 like Figure 1 The diagram shown is a flowchart of a multimodal emotional hallucination assessment and mitigation method provided by an embodiment of the present invention, which mainly includes the following steps: Step 1: Input processing.

[0021] In this embodiment of the invention, reasoning is performed on the input multimodal data based on a multimodal large language model to generate original sentiment interpretation text containing multiple sentences.

[0022] Step 2: Multidimensional assessment.

[0023] In this embodiment of the invention, each sentence is decomposed into the smallest semantic unit, and each smallest semantic unit is classified into a predetermined sentiment analysis dimension; wherein, the sentiment analysis dimension includes a perception layer and a reasoning layer dimension.

[0024] In this embodiment of the invention, each sentence in the sentiment interpretation text is decomposed into the smallest semantic unit, and each smallest semantic unit is classified into a predetermined sentiment analysis dimension, including: For the a sentence It is decomposed into the smallest semantic unit through atomization, and is represented as: ; Where C represents a set of minimal semantic units obtained from the decomposition. , Represents the i-th smallest semantic unit, i=1,2,...,n, where n is the number of smallest semantic units; This is an atomization function; Map each smallest semantic unit to a predetermined sentiment analysis dimension s. , It is a set of sentiment analysis dimensions composed of the perception layer and the reasoning layer.

[0025] For example, the set of sentiment analysis dimensions can include six dimensions: facial expression, behavior, voice, intuition, logic, and conclusion; among them, facial expression, behavior, and voice belong to the perception layer, while intuition, logic, and conclusion belong to the reasoning layer.

[0026] Step 3, Intelligent Agent Routing.

[0027] In this embodiment of the invention, an intelligent agent evaluation mechanism is introduced. Based on the sentiment analysis dimension of each minimum semantic unit, an evaluation model is matched according to the task routing strategy. The evaluation model determines whether the corresponding minimum semantic unit is a hallucination. If so, the hallucination word element in the minimum semantic unit is located, and the minimum semantic unit belonging to the hallucination and the position of the corresponding hallucination word element are stored in the hallucination archive. Based on whether all minimum semantic units are hallucinations, the hallucination rate of each sentence is calculated.

[0028] In this embodiment of the invention, the step of determining whether the corresponding smallest semantic unit is a hallucination by evaluating the model includes: the evaluation model includes a model selection function, a modality selection function, and a cue constructor; the model selection function and the modality selection function are used to select the corresponding model and modality according to the sentiment analysis dimension, respectively; the cue constructor constructs cue text according to the selected model and modality, and determines whether the corresponding smallest semantic unit is a hallucination by using the selected model.

[0029] In this embodiment of the invention, the calculation of the hallucination rate of each sentence based on whether all the smallest semantic units are hallucinations is expressed as follows: ; Where n is the number of the smallest semantic units. The evaluation result for whether the i-th smallest semantic unit is a hallucination. This indicates that the i-th smallest semantic unit is an illusion; It is an exponential function; if the condition in the square brackets is met, the output is 1, otherwise the output is 0; EHR is the illusion rate of each sentence.

[0030] Step 4: Memory Reasoning.

[0031] In this embodiment of the invention, for each sentence in the original sentiment interpretation text, if the hallucination rate is not higher than a threshold, it is directly taken as the optimal result; if the hallucination rate is higher than the threshold, K rounds of iterative correction are performed. During the iterative correction process, information in the hallucination archive is referenced to suppress hallucination lexical terms. For example, if the hallucination rate is greater than 0, K rounds of iterative correction are performed.

[0032] For the a sentence The original score vector output in the k-th iteration is It corresponds to the sentence Each word in the vector represents a score of the illusion; these are then corrected to obtain a corrected score vector. , is represented as: ; in, This is an adjustment coefficient used to control the intensity of the penalty. To The probability distribution obtained after processing with the Softmax function, which is a normalized exponential function, and ⊙ represents element-wise multiplication; This represents the original score vector corresponding to the illusionary morpheme in the previous iteration. When k=1, For the first a sentence The original score vector corresponding to the hallucination word unit is decomposed into the smallest semantic unit after each round of iteration correction. After evaluation by the evaluation model, the hallucination archive is updated according to the evaluation results. The corresponding original score vector is determined according to the smallest semantic unit recorded in the hallucination archive and the position of the corresponding hallucination word unit.

[0033] Step 5: Output the results.

[0034] In this embodiment of the invention, for sentences that undergo K rounds of iterative correction, the sentence with the lowest illusion rate is selected as the optimal result from the K rounds of iterations. The corrected sentiment interpretation text is composed based on the optimal results of all sentences and stored in the anchor archive. The final sentiment interpretation text is obtained by re-inferring based on the multimodal large language model combined with the input multimodal data and the anchor archive.

[0035] For the a sentence The sentence obtained from the k-th iteration is denoted as The iteration with the lowest hallucination rate is determined by the following formula. : ; in, Sentence The illusion rate is used to output the iteration round corresponding to the minimum illusion rate; Then the iteration rounds Corresponding sentence This is the optimal result; Finally, the revised sentiment interpretation text is constructed based on the optimal results of all sentences. Let L be the total number of sentences in the original sentiment interpretation text, then the revised sentiment interpretation text is... , The revised emotional interpretation text Save to anchor point archive (MA): .

[0036] Then, based on the multimodal large language model combined with the input multimodal data and anchor point archive, inference is performed again, where the first... a sentence The reasoning process can be represented as follows: ; in, It is a multimodal large language model.

[0037] The above-mentioned solution provided by the embodiments of the present invention has the following main advantages: First, in terms of emotional hallucination modeling, by introducing psychological structured analysis methods, a fine characterization of complex emotional reasoning processes is achieved, enabling the system to distinguish different levels of error sources, thereby improving hallucination recognition capabilities. Second, in terms of generation control, the present invention embeds an evaluation mechanism into the generation process for the first time, realizing a shift from "post-correction" to "process control," effectively avoiding the problem of error accumulation. Overall, the present invention can significantly reduce the hallucination rate of multimodal large language models in emotional understanding tasks and improve the authenticity and consistency of generated content. At the same time, due to the adoption of semantic atomization and multi-agent collaborative mechanisms, it has good scalability and can flexibly adapt to different types of models and task scenarios. In addition, the fine-grained control method based on logit (raw score) achieves precise suppression of erroneous information while maintaining the original semantic expression ability, without significantly affecting the generation quality.

[0038] To more clearly demonstrate the technical solution and its effects provided by the present invention, the method provided by the embodiments of the present invention will be described in detail below with reference to specific examples.

[0039] I. Overall Overview of the Plan

[0040] This invention provides a multimodal emotional hallucination assessment and mitigation method. This method constructs an assessment-driven closed-loop mechanism to identify and suppress hallucinations in real time during the generation process. The main technologies are as follows: (1) From a psychological perspective, the emotional understanding process is structured into two stages: the perception layer and the reasoning layer, and further refined into six dimensions (facial expression, behavior, voice, intuition, logic, and conclusion), thereby establishing a quantifiable emotional illusion assessment system, which enables the originally difficult-to-define emotional errors to be accurately characterized and analyzed.

[0041] (2) A semantic atomization mechanism was designed to decompose the complex text generated by the model into multiple independent minimum semantic units, so that each sentiment judgment or reasoning step can be evaluated separately. This fine-grained decomposition method solves the problem that traditional overall evaluation is difficult to locate the source of error, and provides a basis for subsequent accurate correction.

[0042] (3) An intelligent agent evaluation mechanism is introduced, which matches the most suitable evaluation model for different types of semantic units through task routing strategy, realizes multi-expert collaborative judgment, and thus significantly improves the accuracy of hallucination recognition. The evaluation results not only give a judgment on whether hallucination exists, but also can locate specific erroneous words (i.e., hallucination words), providing a direct basis for generation and regulation.

[0043] (4) In terms of generation control, this invention proposes a dynamic correction method based on the logit (score) level. By applying probability penalties to historical illusionary lexical units, the model is gradually guided away from the erroneous path during the generation process. Compared with traditional methods based on prompts or resampling, this invention can provide finer control over the model output at a more fundamental level.

[0044] II. Detailed introduction of the plan.

[0045] The core of this invention lies in introducing a closed-loop mechanism of "evaluation-feedback-correction-memory" during the generation process, thereby avoiding the problem of error accumulation in traditional post-processing methods. The overall technical solution of this invention can be summarized as follows: First, the emotional reasoning content (emotional explanation text) generated by the model is structurally decomposed; second, potential illusions are identified through a multi-dimensional evaluation mechanism; third, a memory-based reasoning mechanism is used to gradually intervene in the generation process, ultimately achieving a highly reliable emotional understanding result output. The entire process mainly includes: input processing, multi-dimensional evaluation, agent routing, memory reasoning, and result output, such as... Figure 2 As shown below. Each step will be described in detail below.

[0046] 1. Input processing.

[0047] In this embodiment of the invention, the input is multimodal data. ,in, Indicates video information, Represents audio information. Representing textual information, multimodal data As basic perceptual information, it enters the multimodal large language model, which then gradually generates sentiment interpretation text through an autoregressive generative approach. ,in This indicates the total number of sentences generated. Indicates the first One sentence. In the generation process, multimodal large language models output word-level probability distributions at each step. ,in, Let be a random variable, representing the t-th word to be generated. Let represent the first t generated words. This probability distribution reflects the likelihood of generating each word in the current generation state. The key point of this invention is that it does not directly accept the generation result, but introduces an evaluation and correction mechanism during the generation process.

[0048] 2. Multidimensional assessment.

[0049] In another embodiment of this invention, a multidimensional emotional illusion assessment system is proposed, which breaks down the emotional understanding process into six sub-dimensions: Expression, Action, Audio, Instinct, Logic, and Conclusion. The first three belong to the perception layer, and the latter three belong to the reasoning layer. This applies to emotional interpretation texts. Each sentence in the text is first decomposed into a set of minimal semantic units through atomization. Each term represents a unique minimal semantic unit, where n is the number of minimal semantic units, such as "voice trembling" or "showing anger". a sentence This process can be formally represented as ,in This represents the atomic processing function. Furthermore, for each smallest semantic unit, it is further mapped to a specific evaluation dimension. ,in It is a set of six dimensions (expression, behavior, voice, intuition, logic, and conclusion).

[0050] 3. Intelligent agent routing.

[0051] For each semantic unit, based on its corresponding dimension Selecting the appropriate evaluation model (strategy) comprises three components: model selection function. Modal selection function And hints about constructors For example, when When it is an "expression", the visual model is preferred and the video modality is called; when When the input is "speech", the audio model is invoked. The evaluation function is defined as follows: ,in, The evaluation model is an intelligent agent based on a large language model. The above formula means the evaluation model. For the smallest semantic unit Selecting a strategy on dimension s An assessment will be conducted to determine whether it is a hallucination. ,when This indicates that the semantic unit is a hallucination. This indicates credibility.

[0052] Furthermore, the formula for calculating the illusion rate of each sentence is defined as follows: Where n is the number of the smallest semantic units, This is the indicator function. The aforementioned illusion rate can be further refined into sub-indicators for each dimension and an overall indicator for subsequent optimization decisions.

[0053] 4. Memory reasoning.

[0054] After obtaining the evaluation results, if a certain sentence A value of 0 indicates no hallucination has occurred and no correction is needed; a value greater than 0 initiates the reasoning and correction phase. This invention proposes a generation and control mechanism based on hierarchical memory. This mechanism comprises two key storage structures: a hallucination archive. Archive with anchor points . Used to record semantic units identified as hallucinations and their corresponding lexical positions. Used to store verified, highly reliable sentence results. For the first... a sentence It will undergo K rounds of iterative correction, with each round adjusting the current generation probability based on the hallucination information identified in the previous round.

[0055] In its specific implementation, let the logits generated in the k-th round be... The corrected formula for calculating logits is: ,in This is an adjustment coefficient used to control the intensity of the penalty; its value range is usually [10, 100], with a preferred value of 50. To The probability distribution obtained after applying the Softmax function; ⊙ represents element-wise multiplication; This represents the logits (original score vector) corresponding to the hallucination words identified in the previous round. The meaning of this formula is that by applying probability suppression to historical hallucination words, the probability of their occurrence in the current generation is reduced, thereby achieving dynamic avoidance of erroneous paths.

[0056] 5. Output the results.

[0057] For the a sentence After completing K rounds of iterations, the result set is generated. The result with the lowest illusion rate is selected as the optimal result, corresponding to the iteration round. All sentences undergo the above processing: if no correction is needed, it is directly used as the optimal result; otherwise, K rounds of iterative correction are performed to obtain the optimal result, thus obtaining the corrected sentiment interpretation text. Finally, the corrected sentiment interpretation text is stored in the anchor archive. This step ensures that each round of generation is based on verified and reliable content, thereby avoiding the propagation of errors.

[0058] Subsequently, inference is performed again using a multimodal large language model. This time, it no longer relies on the original autoregressive path, but is based on the input multimodal data. and anchor set Regeneration, i.e. This process, known as re-inference based on trusted anchors, has the core advantage of breaking erroneous dependency chains.

[0059] III. Example Introduction.

[0060] 1. This example applies to a video-based emotion understanding task. The input is a video clip containing dialogue, which includes visual, audio, and subtitle information. First, the input data undergoes unified processing: the video frame sequence is used as visual input, the audio stream is converted into audio features, and subtitle text is extracted or recognized, forming complete multimodal input data. Then, this multimodal data is fed into a multimodal large language model, which generates a sentence-by-sentence emotion interpretation of the video content. The result is a natural language description consisting of multiple sentences, each corresponding to a stage of emotion inference.

[0061] After generating each sentence, the result is not output directly. Instead, the sentence is input into the semantic decomposition module for processing. This module's task is to break down the originally continuous natural language description into multiple independent minimal semantic units. Each minimal semantic unit expresses only a single judgment, such as a description of a person's facial expression, a judgment of tone of voice, or a speculation about a psychological state. In this way, the complex emotional reasoning process can be transformed into multiple individually verifiable minimal units.

[0062] Next, each smallest semantic unit is classified into a predefined sentiment analysis dimension, such as belonging to facial expressions, behaviors, or speech, or to intuition, logic, or conclusions within the reasoning layer. After classification, an appropriate evaluation model is assigned to each semantic unit through agent scheduling. For example, for the smallest semantic unit involving facial expressions, a visual understanding model combined with video frames is used for verification; for semantic units involving speech intonation, a speech analysis model is used for judgment; and for semantic units involving logical reasoning, a language model is used for consistency analysis.

[0063] During the evaluation process, each smallest semantic unit receives a credibility rating, and specific words or fragments within that unit that may contain problems are also flagged. All content deemed untrustworthy is recorded in the illusion archive for reference in subsequent generation processes. Simultaneously, the overall reliability of the current sentence is assessed. If significant illusions are detected (i.e., the illusion rate is greater than 0), the sentence is not directly accepted as the final result but instead enters a correction phase.

[0064] During the correction phase, the multimodal large language model is invoked again to generate an alternative version of the current sentence. However, information from the illusion archive is referenced during the generation process to suppress problematic content from the previous iteration. For example, if the previous generation incorrectly identified the character's emotion as anger, the probability of anger-related expressions is reduced in the next generation, guiding the model to try other more reasonable emotional interpretations. This process can be repeated multiple times, with each iteration incorporating new evaluation results to continuously optimize the generated content.

[0065] After multiple rounds of generation, the most reliable result from all candidate sentences is selected as the final output of the current sentence, and this sentence is stored in the anchor archive. This anchor archive stores verified, high-quality content as contextual input in subsequent sentence generation, ensuring the overall sentiment inference process is built on a reliable foundation. This process is repeated until all sentences are corrected, and finally, the anchor archive is stored within the corrected sentiment explanation text.

[0066] Subsequently, the multimodal large language model no longer relies solely on the original generation path. Instead, it reorganizes the context based on the input multimodal data and already confirmed credible sentences (anchor archives) to generate the final result. This approach effectively avoids the impact of earlier errors on subsequent results, thereby improving the consistency and accuracy of the overall inference. The process is repeated sentence by sentence until the entire sentiment interpretation text is generated.

[0067] 2. In this example, the analysis of user voice emotions is applied in an intelligent customer service scenario. The input is the user's voice call data, which is transcribed into text by a speech recognition module, while simultaneously extracting voice emotion features. Subsequently, the same process is followed to generate emotion interpretation, decompose semantics, evaluate the agent, and iteratively correct, ultimately outputting an accurate description of the user's emotional state. This result can be used to assist the customer service system in adjusting its strategies, such as identifying whether a user is angry or anxious, thereby adopting a more appropriate response.

[0068] As can be seen from the above embodiments, the present invention is not only applicable to video understanding scenarios, but can also be flexibly applied to various tasks such as speech analysis and human-computer interaction. The key to the entire method lies in the deep integration of the generation process and the evaluation mechanism, continuously optimizing the output results through multiple rounds of feedback, thereby achieving higher reliability and accuracy in complex emotion understanding tasks.

[0069] Through the above description of the embodiments, those skilled in the art can clearly understand that the above embodiments can be implemented by software, or by using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions of the above embodiments can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, mobile hard drive, etc.), including several instructions to cause a computer device (such as a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0070] Example 2 This invention also provides a multimodal emotional hallucination assessment and mitigation system, which is mainly used to implement the methods provided in the foregoing embodiments, such as... Figure 3 As shown, the system mainly includes: The input processing module is used to perform reasoning on multimodal data based on a multimodal large language model and generate original sentiment interpretation text containing multiple sentences. The multidimensional evaluation module is used to decompose each sentence into its smallest semantic unit and classify each smallest semantic unit into a predetermined sentiment analysis dimension; the sentiment analysis dimension includes a perception layer and an inference layer dimension. The intelligent agent routing module is used to match the corresponding evaluation model based on the sentiment analysis dimension of each minimum semantic unit using a task-based routing strategy. The evaluation model determines whether the corresponding minimum semantic unit is a hallucination. When it is determined to be a hallucination, the position of the corresponding minimum semantic unit and the position of the corresponding hallucination word are stored in the hallucination archive. Based on whether all minimum semantic units are hallucinations, the hallucination rate of each sentence is calculated. The memory reasoning module is used to directly take the optimal result for each sentence if the hallucination rate is not higher than the threshold; if the hallucination rate is higher than the threshold, it performs K rounds of iterative correction, and during the iterative correction process, it refers to the information in the hallucination archive to suppress hallucination words. The output module is used to select the sentence with the lowest illusion rate from the K rounds of iterations as the optimal result for sentences that have undergone K rounds of iterations. The optimal results of all sentences are used to form the revised sentiment interpretation text and store it in the anchor archive. Based on the multimodal large language model, the input multimodal data and the anchor archive are combined to perform inference again to obtain the final sentiment interpretation text.

[0071] Since the technical details involved in this system have been described in detail in previous embodiments, they will not be repeated here.

[0072] Those skilled in the art will understand that, for the sake of convenience and brevity, the above-described division of functional modules is used as an example. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the system can be divided into different functional modules to complete all or part of the functions described above.

[0073] Example 3 The present invention also provides a processing device, such as Figure 4 As shown, it mainly includes: one or more processors; a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the method provided in the foregoing embodiments.

[0074] Furthermore, the processing device also includes at least one input device and at least one output device; in the processing device, the processor, memory, input device, and output device are connected via a bus.

[0075] In this embodiment of the invention, the specific types of the memory, input device, and output device are not limited; for example: Input devices can be touchscreens, image acquisition devices, physical buttons, or mice, etc. The output device can be a display terminal; The memory can be random access memory (RAM) or non-volatile memory, such as disk storage.

[0076] Example 4 The present invention also provides a readable storage medium storing a computer program that, when executed by a processor, implements the method provided in the foregoing embodiments.

[0077] In this embodiment of the invention, the readable storage medium is a computer-readable storage medium and can be disposed in the aforementioned processing device, for example, as a memory in the processing device. Furthermore, the readable storage medium can also be any medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0078] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims. The information disclosed in the background section is intended only to enhance the understanding of the overall background technology of the present invention and should not be construed as an admission or implication in any way that such information constitutes prior art known to those skilled in the art.

Claims

1. A method for assessing and alleviating multimodal emotional hallucinations, characterized in that, include: Inference is performed on multimodal data based on a multimodal large language model to generate original sentiment interpretation text containing multiple sentences; Each sentence is decomposed into its smallest semantic unit, and each smallest semantic unit is classified into a predetermined sentiment analysis dimension; the sentiment analysis dimension includes a perception layer and a reasoning layer dimension. Based on the sentiment analysis dimension of each minimum semantic unit, an evaluation model is matched according to the task routing strategy. The evaluation model determines whether the corresponding minimum semantic unit is a hallucination. When it is determined to be a hallucination, the position of the corresponding minimum semantic unit and the position of the corresponding hallucination word are stored in the hallucination archive. Based on whether all minimum semantic units are hallucinations, the hallucination rate of each sentence is calculated. For each sentence, if the hallucination rate is not higher than the threshold, it is directly taken as the optimal result; if the hallucination rate is higher than the threshold, K rounds of iterative correction are performed, and the hallucination words are suppressed by referring to the information in the hallucination archive during the iterative correction process. For sentences undergoing K rounds of iteration for correction, the sentence with the lowest illusion rate is selected as the optimal result from the K rounds of iterations. The corrected sentiment interpretation text is then constructed based on the optimal results of all sentences and stored in the anchor archive. The final sentiment interpretation text is obtained by re-inferring based on the multimodal large language model combined with the input multimodal data and the anchor archive.

2. The method for assessing and alleviating multimodal emotional hallucinations according to claim 1, characterized in that, The process of decomposing each sentence into its smallest semantic unit and classifying each smallest semantic unit into a predetermined sentiment analysis dimension includes: For the a sentence It is decomposed into the smallest semantic unit through atomization, and is represented as: ； Where C represents a set of minimal semantic units obtained from the decomposition. , Represents the i-th smallest semantic unit, i=1,2,...,n, where n is the number of smallest semantic units; This is an atomization function; Map each smallest semantic unit to a predetermined sentiment analysis dimension s. , It is a set of sentiment analysis dimensions composed of the perception layer and the reasoning layer.

3. The method for assessing and alleviating multimodal emotional hallucinations according to claim 1, characterized in that, The step of determining whether the corresponding smallest semantic unit is a hallucination through the evaluation model includes: The evaluation model includes a model selection function, a modality selection function, and a cue constructor. The model selection function and modality selection function are used to select the corresponding model and modality based on the sentiment analysis dimension, respectively; the cue constructor constructs cue text based on the selected model and modality, and determines whether the corresponding smallest semantic unit is a hallucination based on the selected model.

4. The method for assessing and alleviating multimodal emotional hallucinations according to claim 1, characterized in that, The calculation of the hallucination rate for each sentence based on whether all the smallest semantic units are hallucinations is expressed as follows: ； Where n is the number of the smallest semantic units. The evaluation result for whether the i-th smallest semantic unit is a hallucination. This indicates that the i-th smallest semantic unit is an illusion; It is an exponential function; if the condition in the square brackets is met, the output is 1, otherwise the output is 0; EHR is the illusion rate of each sentence.

5. The method for assessing and alleviating multimodal emotional hallucinations according to claim 1, characterized in that, The K-round iterative correction includes: For the a sentence The original score vector output in the k-th iteration is Correct it to obtain the corrected score vector. , is represented as: ； in, This is an adjustment coefficient used to control the intensity of the penalty. To The probability distribution obtained after processing with the Softmax function, which is a normalized exponential function, and ⊙ represents element-wise multiplication; This represents the original score vector corresponding to the illusionary word element in the previous iteration. When k=1, For the first a sentence The original score vector corresponding to the hallucination word unit is decomposed into the smallest semantic unit after each round of iteration correction. After evaluation by the evaluation model, the hallucination archive is updated according to the evaluation results. The corresponding original score vector is determined according to the smallest semantic unit recorded in the hallucination archive and the position of the corresponding hallucination word unit.

6. A method for assessing and alleviating multimodal emotional hallucinations according to claim 1 or 5, characterized in that, The process of selecting the sentence with the lowest illusion rate from the K-round iteration results as the optimal result for the sentence undergoing K-round iteration correction includes: For the a sentence The sentence obtained from the k-th iteration is denoted as The iteration with the lowest hallucination rate is determined by the following formula. : ； in, Sentence The illusion rate is used to output the iteration round corresponding to the minimum illusion rate; Then the iteration rounds Corresponding sentence This is the optimal result.

7. The method for assessing and alleviating multimodal emotional hallucinations according to claim 6, characterized in that, The process of re-inferring based on a multimodal large language model, combined with the input multimodal data and anchor archives, yields the final sentiment interpretation text, including: For the a sentence The reasoning process can be represented as follows: ； in, For a multimodal large language model, X represents multimodal data, and MA represents anchor archives.

8. A multimodal emotional hallucination assessment and mitigation system, characterized in that, To implement the method according to any one of claims 1 to 7, comprising: The input processing module is used to perform reasoning on multimodal data based on a multimodal large language model and generate original sentiment interpretation text containing multiple sentences. The multidimensional evaluation module is used to decompose each sentence into its smallest semantic unit and classify each smallest semantic unit into a predetermined sentiment analysis dimension; the sentiment analysis dimension includes a perception layer and an inference layer dimension. The intelligent agent routing module is used to match the corresponding evaluation model based on the sentiment analysis dimension of each minimum semantic unit using a task-based routing strategy. The evaluation model determines whether the corresponding minimum semantic unit is a hallucination. When it is determined to be a hallucination, the position of the corresponding minimum semantic unit and the position of the corresponding hallucination word are stored in the hallucination archive. Based on whether all minimum semantic units are hallucinations, the hallucination rate of each sentence is calculated. The memory reasoning module is used to directly take the optimal result for each sentence if the hallucination rate is not higher than the threshold; if the hallucination rate is higher than the threshold, it performs K rounds of iterative correction, and during the iterative correction process, it refers to the information in the hallucination archive to suppress hallucination words. The output module is used to select the sentence with the lowest illusion rate from the K rounds of iterations as the optimal result for sentences that have undergone K rounds of iterations. The optimal results of all sentences are used to form the revised sentiment interpretation text and store it in the anchor archive. Based on the multimodal large language model, the input multimodal data and the anchor archive are combined to perform inference again to obtain the final sentiment interpretation text.

9. A processing device, characterized in that, include: One or more processors; Memory, used to store one or more programs; Wherein, when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method as described in any one of claims 1 to 7.

10. A readable storage medium storing a computer program, characterized in that, When a computer program is executed by a processor, it implements the method as described in any one of claims 1 to 7.

Images