A multi-modal representation method and system based on implicit preference learning
By employing implicit preference learning, this method utilizes a preference learning model for random segment extraction and implicit preference labeling, combined with reparameterization of the reward function. This achieves efficient resource utilization and trajectory-level language-visual alignment in multimodal representation learning, solving the problems of high resource consumption and low efficiency in existing methods. It is suitable for training large-scale visual language models.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- TSINGHUA UNIVERSITY
- Filing Date
- 2024-02-26
- Publication Date
- 2026-07-28
AI Technical Summary
Existing methods are resource-intensive and inefficient in multimodal representation learning, making it difficult to achieve effective language-visual alignment at the trajectory level. Furthermore, existing methods often sacrifice local or global information to achieve other goals.
By employing an implicit preference learning-based approach, a preference learning model is used for random segment extraction and implicit preference labeling. Combined with reparameterization of the reward function, the alignment of image sequences with text representations is achieved.
It reduces the consumption of learning resources, improves efficiency, solves the problem of trajectory-level representation alignment, provides a principled method for general representation and reward learning, and can be extended to the training of large visual language models.
Smart Images
Figure CN118365980B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence model training, and in particular to a multimodal representation method and system based on implicit preference learning. Background Technology
[0002] Achieving general decision-making has always been the ultimate goal of artificial intelligence research, aiming to build autonomous robotic agents that can understand their environment from raw visual input and perform tasks described in natural language. To pursue such agents, a solution is to train large models end-to-end using action annotations on a wide range of in-domain data. However, this typically creates a huge demand for training data. On the other hand, pre-trained scalable and generalizable visual-language representations can leverage inexpensive out-of-domain data without action annotations, such as massive amounts of human-operation videos, to combat the scarcity of domain-specific data and facilitate effective downstream policy learning. While these general videos may not directly guide specific downstream representations or tasks, their extensive state coverage can greatly enhance generalization. Furthermore, these videos contain interactive dynamics about how objects interact with their surroundings and how task progress is semantically represented across different instructions. Therefore, they have great potential for execution of diverse downstream tasks, eliminating the dependence on intensive, domain-specific data collection.
[0003] Multimodal pre-training has become an effective strategy for achieving the three-pronged goal of representation learning in autonomous robots: 1) extracting local and global task progress information; 2) enhancing the temporal consistency of visual representations; and 3) capturing linguistic information at the trajectory level. Most existing methods achieve these goals through individual objectives, which often result in suboptimal solutions.
[0004] A prerequisite for learning effective multimodal joint representations is trajectory-level language-visual representation alignment, which refers to aligning language with image sequences rather than individual frames, as image sequences (unlike static frames) inherently capture the agent's behavior in fulfilling a given instruction. Focusing on image sequences of varying lengths is crucial, as shorter sequences reveal detailed local transition information, while longer sequences provide semantic information about global task progress. Achieving the appropriate balance between local and global views can be highly complex, but it determines the overall quality of the task representation. Existing methods either tend to focus on short, fixed-length sequences, thus lacking global context, or only consider long-term perspectives that ignore local transitions. Furthermore, they often rely on separate objectives to capture task progress, achieve temporal consistency, and achieve language-visual alignment, balancing these objectives through difficult-to-tune hyperparameters, often at the expense of others. Therefore, developing a general solution that addresses these multifaceted challenges in a unified manner becomes essential. Summary of the Invention
[0005] This invention provides a multimodal representation method and system based on implicit preference learning to solve the problems of high resource consumption and low efficiency in the policy learning of existing downstream models.
[0006] This invention provides a multimodal representation method based on implicit preference learning, comprising:
[0007] Acquire raw video images and text representations;
[0008] The original image sequence of the original video image is extracted, and random segments are extracted from the original image sequence through a preset preference learning model. The extracted segments are implicitly labeled with preferences and the preference reward function is reparameterized to determine the optimal image sequence.
[0009] Align the optimal image sequence with the text representation to complete the matching of the semantics of the text representation with the original video image.
[0010] According to the multimodal representation method based on implicit preference learning provided by the present invention, the acquisition of original video images and text representations specifically includes:
[0011] Obtain multiple raw video images from an existing video library;
[0012] While acquiring multiple raw video images, the text representations of the corresponding language instructions are extracted.
[0013] According to the multimodal representation method based on implicit preference learning provided by the present invention, the step of extracting the original image sequence of the original video image and randomly extracting segments from the original image sequence through a preset preference learning model specifically includes:
[0014] Convert the original video images into an original image sequence;
[0015] Randomly extract image segments based on the original image sequence;
[0016] During the extraction process, images are randomly extracted from the middle and tail of the original image sequence, respectively.
[0017] According to the multimodal representation method based on implicit preference learning provided by the present invention, the implicit preference annotation of the extracted segments specifically includes:
[0018] Find the text annotations of the original video images based on the extracted segments;
[0019] Preference labels are determined based on the text annotations, and implicit preference labeling is performed on the extracted fragments.
[0020] According to the multimodal representation method based on implicit preference learning provided by the present invention, the reparameterization of the preference reward function specifically includes:
[0021] Calculate the first absolute distance between the head static image and the text representation and the second absolute distance between the tail static image and the text representation in the extracted segment;
[0022] The maximum difference between the first absolute distance and the second absolute distance is determined as the reward function over the absolute distance.
[0023] According to the multimodal representation method based on implicit preference learning provided by the present invention, the reparameterization of the preference reward function further includes:
[0024] The tail image and head image in the extracted segment are compared in terms of transfer direction, and the transfer difference is calculated.
[0025] The transfer difference is compared with the transfer direction difference between the text representation;
[0026] When the transition difference and the text representation have similar transition directions, the transition difference is determined as the reward function in the transition direction, thus narrowing the distance between the extracted fragment and the text representation.
[0027] The present invention also provides a multimodal representation system based on implicit preference learning, the system comprising:
[0028] The data acquisition module is used to acquire raw video images and text representations;
[0029] The image processing module extracts the original image sequence of the original video image, randomly extracts segments from the original image sequence using a preset preference learning model, implicitly labels the extracted segments with preferences, reparameterizes the preference reward function, and determines the optimal image sequence.
[0030] The alignment module is used to align the optimal image sequence with the text representation, thereby matching the semantics of the text representation with the original video image.
[0031] 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 multimodal representation method based on implicit preference learning as described above.
[0032] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the multimodal representation method based on implicit preference learning as described above.
[0033] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the multimodal representation method based on implicit preference learning as described above.
[0034] This invention provides a multimodal representation method and system based on implicit preference learning. By introducing intelligent design, it achieves alignment between original video and text representations based on implicit preference annotation, random segment sampling, and reparameterization of the reward function in the embedding space. The preference learning model can be quickly extended to visual language representation learning, generating training data suitable for current needs, reducing learning resource consumption, and improving efficiency. The derived decision framework effectively solves the core challenges in trajectory-level representation alignment, providing a principled method for general representation and reward learning, and can further expand the decision to train large-scale visual language models. Attached Figure Description
[0035] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0036] Figure 1 This is a flowchart illustrating a multimodal representation method based on implicit preference learning provided by the present invention.
[0037] Figure 2 This is a schematic diagram of the module connection of a multimodal representation system based on implicit preference learning provided by the present invention;
[0038] Figure 3 This is a schematic diagram of the structure of the electronic device provided by the present invention.
[0039] Figure label:
[0040] 110: Data acquisition module; 120: Image processing module; 130: Alignment module;
[0041] 310: Processor; 320: Communication interface; 330: Memory; 340: Communication bus Detailed Implementation
[0042] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.
[0043] The following is combined with Figure 1 This invention describes a multimodal representation method based on implicit preference learning, comprising:
[0044] S100: Acquire raw video images and text representations;
[0045] In this invention, multiple original video images are obtained from an existing video library;
[0046] While acquiring multiple raw video images, the text representations of the corresponding language instructions are extracted.
[0047] This invention proposes a general unified objective to simultaneously extract meaningful task progress information from image sequences and seamlessly align it with language instructions. Through implicit preferences, where image sequences can be better aligned with their corresponding language instructions, this is transformed into representation learning through appropriate reward reparameterization. A multimodal learning framework, transformed through preference learning, can extract local and global task progress features, enforce temporal consistency through implicit temporal contrastive learning, and ensure the instruction basis of the image sequence through multimodal joint encoding, providing a general solution for unified representation and reward learning.
[0048] S200. Extract the original image sequence of the original video image, randomly extract segments from the original image sequence using a preset preference learning model, implicitly label the extracted segments with preferences and reparameterize the preference reward function to determine the optimal image sequence.
[0049] Specifically, the original image sequence of the original video image is extracted, and random segments are extracted from the original image sequence using a preset preference learning model, specifically including:
[0050] Convert the original video images into an original image sequence;
[0051] Randomly extract image segments based on the original image sequence;
[0052] During the extraction process, images are randomly extracted from the middle and tail of the original image sequence, respectively.
[0053] In practical applications, language is aligned with image sequences rather than individual frames because image sequences (unlike static frames) inherently capture the agent's behavior in fulfilling a given instruction. Focusing on image sequences of varying lengths is crucial, as shorter sequences reveal detailed local transition information, while longer sequences provide global task progress semantics. Finding the appropriate balance between local and global views can be very complex, but it determines the overall quality of the task representation. Existing methods either tend to focus on short, fixed-length sequences, thus lacking global context, or only consider long-term perspectives that ignore local transitions, making it difficult to focus on overall changes.
[0054] Therefore, this invention uses a multimodal representation learning framework to consider a language-conditional Markov decision process (MDP).
[0055]
[0056] It consists of state space and action space. This is a system dynamics transition model, where γ∈(0,1) is the discount factor. It is assumed that the state space is... Defined on RGB images. It is the language instruction space, which specifies the expected task behavior. It is a language-conditional reward function. The goal is to learn a policy. It can solve language Any task described in the text.
[0057] Suppose a video dataset Each of them The dataset consists of raw RGB images, with each video v accompanied by a descriptive language annotation l. h represents the video length. The dataset D contains various out-of-domain datasets, but no action labels are available. Our goal is to train a visual encoder. and a language encoder The original image o and language l are mapped to a shared k-dimensional visual-language representation space using dataset D. Then, the policy π (φ(o), ψ(l)) can use these representations as input for policy learning through imitation learning or reinforcement learning.
[0058] The Bradley-Terry (BT) model, a pre-defined preference learning model, provides a natural way to convey human preferences, requiring humans to rank fragments. The BT model then optimizes the reward to maximize the result calculated in Equation 1, thus fitting human preferences.
[0059]
[0060] whereσ=(o n.... o n+m. ) represents a segment of length m. σ + >σ - σ in section + Priority over σ - , for σ + >σ - The preference varies with σ + The total reward grows exponentially with σ - Decreasing total reward
[0061] The BT model in Formula 1 is essentially a trajectory-level contrastive learning model that compares different trajectory segments and favors the positive ones. This latent equivalence incentive leverages the BT model for representation learning. By reparameterizing the reward with visual and linguistic representations φ(o) and ψ(l), the original BT model is extended to its linguistically conditional version:
[0062]
[0063] By comparing segments under different instructions using Formula 2, semantic features related to the task are extracted from image sequence segments. This provides a new possibility for achieving trajectory-level language foundation by adapting the BT model. Due to the presence of directions that need optimization in Formula 2, (1) the cost of collecting preference labels is very high, and existing video language datasets do not have clear preference annotations; (2) it is difficult to control the appropriate segment length, because excessively long segments will cause a lot of computation and lose local transfer details, while short segments may lose global task information; (3) for multimodal representation learning, the best reward function in a specific parameterized form is difficult to determine.
[0064] In this invention, implicit preference labeling is performed on the extracted fragments, specifically including:
[0065] Find the text annotations of the original video images based on the extracted segments;
[0066] Preference labels are determined based on the text annotations, and implicit preference labeling is performed on the extracted fragments.
[0067] Although there are no explicit preference labels in existing datasets, language instructions are a natural way for humans to convey intent. It is reasonable to assume that a video is more likely to be preferred by human beings for language instructions that match it than for other non-matching instructions. Existing video language datasets already contain a wealth of such implicit preferences that have not been fully utilized, thus eliminating the need for additional preference annotations.
[0068] When sampling random segments, it is difficult to quickly determine a specific judgment length that can cover both local and global aspects. In this invention, a starting image is randomly selected from a video image sequence, and then the remaining images arranged in sequence are randomly obtained. By repeating the above operation, different segment lengths can be selected, so that the preference model can capture both local transition information in short segments and global task progress in longer segments.
[0069] The preference reward function is reparameterized, specifically including:
[0070] Calculate the first absolute distance between the head static image and the text representation and the second absolute distance between the tail static image and the text representation in the extracted segment;
[0071] The maximum difference between the first absolute distance and the second absolute distance is determined as the reward function over the absolute distance.
[0072] Also includes:
[0073] The tail image and head image in the extracted segment are compared in terms of transfer direction, and the transfer difference is calculated.
[0074] The transfer difference is compared with the transfer direction difference between the text representation;
[0075] When the transition difference and the text representation have similar transition directions, the transition difference is determined as the reward function in the transition direction, thus narrowing the distance between the extracted fragment and the text representation.
[0076] Specifically, the reward function for absolute distance is defined as the language embedding distance difference in Formula 3;
[0077] r P (o t ,o t+1 ,φ,ψ,l):=S(φ(o t+1 ),ψ(l))-S(φ(o t ),ψ(l)) (3)
[0078] S represents cosine similarity. This reward behavior is similar to distance-based reward shaping, which is beneficial for policy learning. Substituting Equation 3 into Equation 2, the total reward for segment σ is:
[0079] ∑ t r P (o t ,o t+1 ,φ,ψ,l)=ΔS(φ n+m ,φ n ;ψ(l)) (4)
[0080] In the formula ΔS(φ n+m ,φn ;ψ(l))=S(φ(o t+1 ),ψ(l))-S(φ(o t ),ψ(l)) represents the distance change, and then, based on r, we obtain P The BT model is instantiated as follows:
[0081]
[0082] Maximize Formula 5 will attract arrive At the same time from Repulsion ψ(l) + This results in a sensory deviation that makes the final image more consistent with the instructions.
[0083] For the reward function along the transition direction, the language represents a transition direction rather than the final image, because instructions are more likely to describe dynamic behavior than a static image. To achieve this, the total reward in Equation 4 is modified to Equation 6, thus obtaining the representation of the reward function along the transition direction:
[0084] ∑ t r T (o t ,o t+1 ,φ,ψ,l)=S(φ(o n+m )-φ(o n ),ψ(l)) (6)
[0085]
[0086] Another representation learning objective is obtained, and the maximization formula will and text embedding vector ψ(l + Align the transfer direction.
[0087] S300. Align the optimal image sequence with the text representation to complete the matching of the text representation semantics with the original video image.
[0088] The preference learning model was pre-trained using large-scale data, such as human video datasets or large datasets of robot manipulation videos, and extensive experiments were conducted in simulated and real robotic environments. Specifically, the study investigated how the learned representations can support effective downstream policy learning using methods such as linguistic conditional behavior cloning, and how they can provide general linguistic conditional rewards that describe the desired direction of task progress.
[0089] This invention provides a multimodal representation method based on implicit preference learning. By introducing intelligent design, it achieves alignment between original video and text representations through implicit preference annotation, random segment sampling, and reparameterization of the reward function in the embedding space. The preference learning model can be rapidly extended to visual language representation learning, generating training data suitable for current needs, reducing learning resource consumption, and improving efficiency. The derived decision framework effectively solves the core challenges in trajectory-level representation alignment, providing a principled method for general representation and reward learning, and can further expand the decision-making process to train large-scale visual language models.
[0090] refer to Figure 2 The present invention also discloses a multimodal representation system based on implicit preference learning, the system comprising:
[0091] Data acquisition module 110 is used to acquire raw video images and text representations;
[0092] The image processing module 120 extracts the original image sequence of the original video image, randomly extracts segments from the original image sequence through a preset preference learning model, implicitly labels the extracted segments with preferences and reparameterizes the preference reward function to determine the optimal image sequence.
[0093] Alignment module 130 is used to align the optimal image sequence with the text representation to complete the matching of the semantics of the text representation with the original video image.
[0094] The data acquisition module retrieves multiple raw video images from the existing video library.
[0095] While acquiring multiple raw video images, the text representations of the corresponding language instructions are extracted.
[0096] The image processing module converts the original video images into an original image sequence;
[0097] Randomly extract image segments based on the original image sequence;
[0098] During the extraction process, images are randomly extracted from the middle and tail of the original image sequence, respectively.
[0099] Implicit preference labeling is applied to the extracted fragments, specifically including:
[0100] Find the text annotations of the original video images based on the extracted segments;
[0101] Preference labels are determined based on the text annotations, and implicit preference labeling is performed on the extracted fragments.
[0102] The preference reward function is reparameterized, specifically including:
[0103] Calculate the first absolute distance between the head static image and the text representation and the second absolute distance between the tail static image and the text representation in the extracted segment;
[0104] The maximum difference between the first absolute distance and the second absolute distance is determined as the reward function over the absolute distance.
[0105] Reparameterizing the preference reward function also includes:
[0106] The tail image and head image in the extracted segment are compared in terms of transfer direction, and the transfer difference is calculated.
[0107] The transfer difference is compared with the transfer direction difference between the text representation;
[0108] When the transition difference and the text representation have similar transition directions, the transition difference is determined as the reward function in the transition direction, thus narrowing the distance between the extracted fragment and the text representation.
[0109] This invention provides a multimodal representation system based on implicit preference learning. By introducing intelligent design, it achieves alignment between original video and text representations through implicit preference annotation, random segment sampling, and reparameterization of the reward function in the embedding space. The preference learning model can be rapidly extended to visual language representation learning, generating training data suitable for current needs, reducing learning resource consumption, and improving efficiency. The derived decision framework effectively solves the core challenges in trajectory-level representation alignment, providing a principled method for general representation and reward learning, and can further expand the decision-making process to train large-scale visual language models.
[0110] Figure 3 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 3 As shown, the electronic device may include a processor 310, a communications interface 320, a memory 330, and a communication bus 340, wherein the processor 310, the communications interface 320, and the memory 330 communicate with each other via the communication bus 340. The processor 310 can call logical instructions in the memory 330 to execute a multimodal representation method based on implicit preference learning, which includes: acquiring original video images and text representations;
[0111] The original image sequence of the original video image is extracted, and random segments are extracted from the original image sequence through a preset preference learning model. The extracted segments are implicitly labeled with preferences and the preference reward function is reparameterized to determine the optimal image sequence.
[0112] Align the optimal image sequence with the text representation to complete the matching of the semantics of the text representation with the original video image.
[0113] Furthermore, the logical instructions in the aforementioned memory 330 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0114] On the other hand, the present invention also provides a computer program product, the computer program product including a computer program, the computer program being able to be stored on a non-transitory computer-readable storage medium, and when the computer program is executed by a processor, the computer is able to execute a multimodal representation method based on implicit preference learning provided by the above methods, the method including: acquiring original video images and text representations;
[0115] The original image sequence of the original video image is extracted, and random segments are extracted from the original image sequence through a preset preference learning model. The extracted segments are implicitly labeled with preferences and the preference reward function is reparameterized to determine the optimal image sequence.
[0116] Align the optimal image sequence with the text representation to complete the matching of the semantics of the text representation with the original video image.
[0117] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform a multimodal representation method based on implicit preference learning provided by the above methods, the method comprising: acquiring original video images and text representations;
[0118] The original image sequence of the original video image is extracted, and random segments are extracted from the original image sequence through a preset preference learning model. The extracted segments are implicitly labeled with preferences and the preference reward function is reparameterized to determine the optimal image sequence.
[0119] Align the optimal image sequence with the text representation to complete the matching of the semantics of the text representation with the original video image.
[0120] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.
[0121] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.
[0122] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A multimodal representation method based on implicit preference learning, characterized in that, include: Acquire raw video images and text representations; The original image sequence of the original video image is extracted, and random segments are extracted from the original image sequence using a preset preference learning model. Implicit preference labeling is performed on the extracted segments, and the preference reward function is reparameterized to determine the optimal image sequence. The preset preference learning model is the Bradley-Terry model, which requires human sorting of the segments. Align the optimal image sequence with the text representation to complete the matching of the semantics of the text representation with the original video image; The step of extracting the original image sequence from the original video image, and randomly extracting segments from the original image sequence using a preset preference learning model, specifically includes: Convert the original video images into an original image sequence; Randomly extract image segments based on the original image sequence; During the extraction process, images are randomly extracted from the middle and tail portions of the original image sequence, respectively. The reparameterization of the preference reward function specifically includes: Calculate the first absolute distance between the head static image and the text representation and the second absolute distance between the tail static image and the text representation in the extracted segment; The maximum difference between the first absolute distance and the second absolute distance is determined as the reward function over the absolute distance; The reparameterization of the preference reward function also includes: The tail image and head image in the extracted segment are compared in terms of transfer direction, and the transfer difference is calculated. The transfer difference is compared with the transfer direction difference between the text representation; When the transition difference and the text representation have similar transition directions, the transition difference is determined as the reward function in the transition direction, thus narrowing the distance between the extracted fragment and the text representation.
2. The multimodal representation method based on implicit preference learning according to claim 1, characterized in that, The acquisition of the original video images and text representations specifically includes: Obtain multiple raw video images from an existing video library; While acquiring multiple raw video images, the text representations of the corresponding language instructions are extracted.
3. The multimodal representation method based on implicit preference learning according to claim 1, characterized in that, The implicit preference labeling of the extracted segments specifically includes: Find the text annotations of the original video images based on the extracted segments; Preference labels are determined based on the text annotations, and implicit preference labeling is performed on the extracted fragments.
4. A multimodal representation system based on implicit preference learning, characterized in that, The system includes: The data acquisition module is used to acquire raw video images and text representations; The image processing module extracts the original image sequence from the original video image, randomly extracts segments from the original image sequence using a preset preference learning model, implicitly labels the extracted segments with preferences, and reparameterizes the preference reward function to determine the optimal image sequence; wherein, the preset preference learning model is the Bradley-Terry model, which requires human sorting of the segments; The alignment module is used to align the optimal image sequence with the text representation, thereby matching the semantics of the text representation with the original video image. The image processing module is further configured to convert the original video image into an original image sequence; randomly extract image segments based on the original image sequence; and randomly extract images from the head, middle, and tail of the original image sequence during the extraction process. The image processing module is further configured to calculate the first absolute distance between the head static image and the text representation and the second absolute distance between the tail static image and the text representation in the extracted segment; and to determine the maximum value of the difference between the first absolute distance and the second absolute distance as the reward function on the absolute distance. The image processing module is further configured to compare the transfer direction of the tail image and the head image in the extracted segment and calculate the transfer difference; compare the transfer difference with the transfer direction difference between the text representation; and, if the transfer difference and the text representation have similar transfer directions, determine the transfer difference as the reward function in the transfer direction to shorten the distance between the extracted segment and the text representation.
5. 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 multimodal representation method based on implicit preference learning as described in any one of claims 1 to 3.
6. 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 multimodal representation method based on implicit preference learning as described in any one of claims 1 to 3.
7. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the multimodal representation method based on implicit preference learning as described in any one of claims 1 to 3.