Dialogue intention analysis method and related equipment
By compressing the history of multi-turn dialogues into a low-dimensional vector and introducing an intent conflict detection algorithm, the problem of discrepancy between the content generated and the user's intent in multi-turn dialogues by traditional AI systems is solved, achieving more accurate user intent recognition and consistency in the generated results.
Patent Information
- Application Number
- CN202511246331.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-02
- Publication Date
- 2025-12-26
AI Technical Summary
Traditional AI systems cannot effectively process multi-turn dialogue history, resulting in deviations between generated content and users' long-term intentions. Furthermore, they cannot identify the priority and dependencies of intentions, leading to inconsistencies in the generated results.
By compressing the multi-turn dialogue history into a low-dimensional vector, the core user intent and operational dependencies are preserved, and an intent conflict detection algorithm is introduced to automatically identify and resolve contradictions between multi-turn instructions.
It improves the accuracy of user intent recognition, enhances the contextual coherence of generative AI in text and image interactions, automatically resolves contradictions between multi-turn commands, and generates results that better meet user expectations.
Smart Images
Figure CN121212331A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the fields of artificial intelligence and natural language processing technology, and in particular to dialogue intent parsing methods and related equipment. Background Technology
[0002] With the development of multimedia and network technologies, there are more and more scenarios where artificial intelligence systems are used to process user commands. However, traditional AI systems can usually only process a single user command, resulting in a deviation between the generated content and the user's long-term intentions.
[0003] Therefore, a new approach is needed to improve the accuracy of identifying user intent. Summary of the Invention
[0004] This application provides a method and related device for parsing dialogue intent, which can effectively utilize contextual information from multi-turn dialogue history to improve the accuracy of recognizing user intent. The technical solution is as follows:
[0005] In a first aspect of this application, a method for parsing dialogue intent is provided, the method comprising:
[0006] Responding to input operations, retrieve the input content;
[0007] By combining the input content and the content of multiple rounds of historical dialogue, a low-dimensional intent vector is parsed and obtained;
[0008] The low-dimensional intent vector is encoded by combining image features, and a generation instruction is output.
[0009] The generation instructions are processed to output the generation result;
[0010] The multi-round historical dialogue content includes multi-round historical input content and multi-round historical generation results.
[0011] This application provides a novel method for parsing dialogue intent. By compressing multi-turn dialogue history into a low-dimensional vector, it preserves the user's core intent and operational dependencies, enhancing the contextual coherence of generative AI in text-image interactions. It also introduces an intent conflict detection algorithm to automatically identify and resolve contradictions between multi-turn instructions.
[0012] In one embodiment, after obtaining the input content, the method further includes:
[0013] Record the operation type and timestamp;
[0014] The operation types include at least: adding, modifying, and deleting.
[0015] In one embodiment, combining the input content and multi-turn historical dialogue content, parsing and obtaining a low-dimensional intent vector includes:
[0016] Calculate information entropy based on search input and multi-turn historical dialogue content;
[0017] Based on the information entropy, high-value dialogue segments are selected;
[0018] Extract key operation nodes from multi-round high-value dialogue segments to generate a key operation chain;
[0019] The key operation chain is compressed into the low-dimensional intent vector.
[0020] In one embodiment, after parsing and passing the low-dimensional intent vector, the method further includes:
[0021] For each key operational node, identify the corresponding intent category and confidence score;
[0022] Priorities are assigned to intent categories based on the operation type and the timestamp.
[0023] The step of encoding the low-dimensional intent vector in conjunction with image features and outputting generation instructions includes:
[0024] The low-dimensional intent vector is encoded by combining the image features and the assigned priority, and the generation instruction is output.
[0025] In one embodiment, the intent category includes at least: object manipulation, attribute adjustment, and style transfer.
[0026] In one embodiment, after filtering high-value dialogue fragments, the method further includes:
[0027] Identify whether there are contradictions among the multi-turn high-value dialogue segments;
[0028] If so, the follow-up questioning process will be initiated, and multiple processing options will be provided;
[0029] In response to a selection operation for any processing option, the processing indicated by the processing option is performed on the multi-turn high-value dialogue segment.
[0030] In a second aspect of the embodiments of this application, an AI system is provided, the system comprising a multimodal input layer, an intent parsing module, a dynamic embedding engine, and a generation module connected in sequence;
[0031] The multimodal input layer is used to respond to the input operation and obtain the input content;
[0032] The intent parsing module is used to combine the input content and the content of multiple rounds of historical dialogue to parse and obtain a low-dimensional intent vector;
[0033] The dynamic embedding engine is used to encode the low-dimensional intent vector in combination with image features and output generation instructions.
[0034] The generation module is used to process the generation instructions and output the generation results;
[0035] The multi-round historical dialogue content includes multi-round historical input content and multi-round historical generation results.
[0036] In a third aspect of the embodiments of this application, an electronic device is provided, including a processor and a memory, wherein the memory stores at least one computer instruction, which is loaded and executed by the processor to perform the method of any one of the first aspects and related embodiments of the first aspect.
[0037] In a fourth aspect of the embodiments of this application, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program, which, when executed by a processor, implements the method as described in any one of the first aspects and related embodiments of the first aspect above.
[0038] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description
[0039] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.
[0040] Figure 1 This is a schematic diagram of the structure of an AI system provided in an embodiment of this application;
[0041] Figure 2 This is a schematic diagram of the structure of an intent parsing module provided in an embodiment of this application;
[0042] Figure 3 This is a schematic diagram of the structure of a dynamic embedding engine provided in an embodiment of this application;
[0043] Figure 4 This is an example diagram illustrating how to determine an intent category according to an embodiment of this application;
[0044] Figure 5 This is a flowchart illustrating a dialogue intent parsing method provided in an embodiment of this application;
[0045] Figure 6This is a flowchart illustrating another dialogue intent parsing method provided in the embodiments of this application. Detailed Implementation
[0046] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.
[0047] Traditional generative AI systems (such as ChatGPT and Midjourney) can typically only process single user commands. When a user issues multiple commands, the AI system can only simply concatenate the dialogue history into a long text input, failing to effectively extract key information from the multi-turn dialogue history and understand the user's long-term intent. This can lead to a significant deviation between the final generated content and the user's long-term intent.
[0048] In multi-turn dialogue interactions, users may issue multiple commands for a single need, but existing systems often cannot identify intent priorities and dependencies, leading to inconsistent generated results.
[0049] In addition, users may issue contradictory commands multiple times, a problem that the existing system cannot effectively solve.
[0050] In view of this, embodiments of this application provide a new method for parsing dialogue intent, which preserves the user's core intent and operational dependencies by compressing the multi-turn dialogue history into a low-dimensional vector; it also introduces an intent conflict detection algorithm to automatically identify and resolve contradictions between multi-turn instructions.
[0051] Next, combined Figures 1 to 4 The structure of the AI system provided in the embodiments of this application will be briefly introduced first. Figure 1 This is a schematic diagram of the structure of an AI system provided in an embodiment of this application.
[0052] like Figure 1 As shown in the embodiments of this application, the AI system may include multiple modules, such as: a multimodal input layer, an intent parsing module, a dynamic embedding engine, and a feedback optimization module.
[0053] The multimodal input layer, intent parsing module, dynamic embedding engine, and feedback optimization module are connected in sequence, and the dynamic embedding engine is also connected to the multimodal input layer.
[0054] The multimodal input layer is used to receive user input operations, which can support multimodal input such as text, images, voice, and gestures. In this application, after receiving user input operations, the multimodal input layer can record the input content, operation type, and timestamp corresponding to the current input operation.
[0055] The user's operation types include at least adding, modifying, and deleting.
[0056] The intent parsing module receives the input content from the multimodal input layer and combines it with the history of multi-turn dialogues to perform intent parsing, thereby obtaining a low-dimensional intent vector, intent category, and confidence score.
[0057] Figure 2 This is a schematic diagram of the structure of an intent parsing module provided in an embodiment of this application.
[0058] Optionally, such as Figure 2 As shown, the intent parsing module may include a dialogue history compression unit and an intent classifier connected in sequence. The dialogue history compression unit is used to combine the input content and multi-turn dialogue history to extract key operation nodes, generate key operation chains, and compress the key operation chains to obtain a low-dimensional intent vector. The intent classifier is used to identify the key operation nodes to determine the intent category corresponding to each key operation node and output the corresponding confidence score.
[0059] The intent categories include at least the following types: object manipulation, attribute adjustment, style transfer, etc.
[0060] It should be understood that the dialogue history includes previous rounds of input and output results.
[0061] It should be noted that a low-dimensional intent vector is a representation that maps high-dimensional data into a low-dimensional dense vector. In this application, it is used to capture the semantic or feature associations of multi-turn dialogue history, facilitating subsequent understanding and processing.
[0062] A confidence score is a quantitative metric that measures the reliability of a model, system, or prediction result. Its core value lies in providing a reference for decision-making and helping people judge the credibility of the results. In this application, a higher confidence score indicates that the intent classifier has correctly classified the intent category, and the higher the credibility; a lower confidence score indicates that the intent classifier has incorrectly classified the intent category, and the lower the credibility.
[0063] The dynamic embedding engine is used to dynamically assign priorities to intent categories based on operation type and timestamp. Then, it encodes the low-dimensional intent vectors based on the priorities and outputs the generated instructions.
[0064] Figure 3This is a schematic diagram of the structure of a dynamic embedding engine provided in an embodiment of this application.
[0065] Optionally, such as Figure 3 As shown, the dynamic embedding engine can include a connected priority allocator and a cross-modal memory network model. The priority allocator is connected to the multimodal input layer and the intent classifier, respectively. The priority allocator is used to dynamically adjust the priority corresponding to the intent category based on the operation type and timestamp (which can also be collectively referred to as user history behavior).
[0066] The cross-modal memory network model is connected to the dialogue history compression unit. The cross-modal memory network model is used to jointly encode low-dimensional intent vectors and image features based on the combination priority to generate context-aware generation instructions.
[0067] It should be understood that Cross-modal Memory Networks (CMNs) are artificial intelligence models that combine multimodal data processing and memory mechanisms, aiming to effectively capture the complex relationships between multi-source heterogeneous data.
[0068] The feedback optimization module is designed with a multimodal reward function. It optimizes the parameters of the cross-modal memory network model through reinforcement learning and user behavior analysis, thereby optimizing the generated instructions.
[0069] In this embodiment of the application, reinforcement learning can be performed based on the PPO algorithm, which stands for Proximal Policy Optimization, to prevent the policy from deviating excessively by limiting the policy update magnitude.
[0070] It should be understood that the multimodal reward function is a function used in reinforcement learning to evaluate the quality of the model's output under multimodal inputs (such as images, text, audio, etc.). It combines information from multiple modalities to calculate a comprehensive reward value to guide the model to learn behaviors that are more in line with expectations.
[0071] The reward signals may include explicit user feedback (such as confirmation or cancellation, user satisfaction rating), semantic consistency between the generated result and the historical intent (such as using CLIP Score to determine), and the reduction rate of operation steps (such as from 5 to 2).
[0072] It should be understood that CLIP Score is a no-reference metric for evaluating image captions, which calculates the semantic similarity between the image and the generated description based on the CLIP model.
[0073] User behavior analysis is used to indicate implicit behavior analysis, where implicit behavior may include user engagement duration, number of user actions, etc.
[0074] Optionally, in Figure 1 The AI system shown may also include a generation module, with a multimodal input layer and a dynamic embedding engine connected to the generation module. The generation module processes the input content according to generation instructions and outputs the generated result.
[0075] Further optionally, the generation module may include a text generation module and an image generation module. The text generation module is used to process the text in the input content according to the generation instructions, and obtain and output the text result; the image generation module is used to process the image in the input content according to the generation instructions, and obtain and output the image result.
[0076] Optionally, in Figure 2 The intent parsing module shown may include an intent conflict detection subunit between the innovation algorithm subunit and the compression algorithm subunit. The intent conflict detection subunit is used to automatically detect intent conflicts in high-value dialogue fragments extracted by the innovation algorithm subunit. When a conflict is detected in a multi-turn dialogue, a follow-up questioning process can be initiated to resolve the conflict.
[0077] Next, combined Figure 5 This application provides a detailed description of the dialogue intent parsing method provided in the embodiments of this application. This method is applied to the above-mentioned... Figure 1 The AI system involved. The method may include steps S110 to S190, which will be described in detail below.
[0078] S110. In response to an input operation, obtain the input content and record the operation type and timestamp.
[0079] Optionally, in response to an input operation, the multimodal input layer can acquire the input content and record the operation type and timestamp.
[0080] The input content supports multimodal input, including text, images, voice, and gestures.
[0081] The operation types must include at least adding, modifying, and deleting.
[0082] S120. Combining the input content and the content of multiple rounds of historical dialogue, analyze and obtain the low-dimensional intent vector, intent category, and confidence score.
[0083] Optionally, the intent parsing module may include a dialogue history compression unit and an intent classifier connected in sequence. The dialogue history compression unit is used to combine the current input content and the historical input content from multiple rounds to extract key operation nodes and generate a key operation chain.
[0084] Optionally, such as Figure 2As shown, the dialogue history compression unit may include a first algorithm subunit and a second algorithm subunit connected in sequence. The first algorithm subunit is connected to the multimodal input layer, and the second algorithm subunit is connected to the intent classifier and the dynamic embedding engine, respectively.
[0085] The first algorithm subunit, also known as the innovative algorithm subunit, is used to combine the input content and the history of multi-turn dialogues to calculate information entropy. Based on this information entropy, it filters high-value dialogue segments and discards low-information dialogue segments or redundant information, such as repetitive instructions.
[0086] It should be noted that information entropy is the theoretical limit of data compression. The higher the entropy value, the more information it contains. The lower the entropy value, the less information it contains and the more redundant content (such as repetitive text or regular images).
[0087] The second algorithm subunit, also known as the compression algorithm subunit, is used to extract key operation nodes from high-value dialogue segments, generate key operation chains, and compress the key operation chains to generate low-dimensional intent vectors.
[0088] Among them, such as Figure 2 As shown, the second algorithm subunit can use a bidirectional GRU and self-attention mechanism to extract key operation nodes in high-value dialogue segments, generate key operation chains, and compress them.
[0089] It should be noted that the Bidirectional Gated Recurrent Unit (GRU) is a recurrent neural network structure that extends the GRU (Gated Recurrent Unit). By simultaneously utilizing the forward information (from left to right) and backward information (from right to left) of the sequence, it can more comprehensively capture the contextual dependencies in the sequence data.
[0090] The self-attention mechanism is used to calculate the "relevance score" of each element (i.e., the key operation node) in the sequence (i.e., the key operation chain involved in this application) with respect to all other elements, and then to perform a weighted summation of the information of other elements based on the score to obtain the "context-enhanced representation" (i.e., the low-dimensional intent vector) of that element.
[0091] It should be understood that the multiple critical operation nodes in the critical operation chain have a sequential order.
[0092] The intent classifier is used to identify the corresponding intent category for each operation node and output the corresponding confidence score.
[0093] The intent categories include at least the following types: object manipulation, attribute adjustment, style transfer, etc.
[0094] For example, such as Figure 4 As shown, the dialogue history compression unit in the intent parsing module extracts the following high-value dialogue fragments from the input content in the first round: "Design a technology company logo with blue as the main color"; in the second round, it extracts "Add a circuit board texture"; and in the third round, it extracts "Blue is too dark, change it to light blue." For these high-value dialogue fragments, the dialogue history compression unit extracts the key operation nodes as "Set the main color," "Add texture," and "Adjust color brightness." Then, a key operation chain can be generated: "Set the main color → Add texture → Adjust color brightness." This key operation chain can then be compressed into a low-dimensional intent vector.
[0095] In addition, for key operation nodes, the intent classifier in the intent parsing module can identify the intent category corresponding to "set main color tone" as "style transfer", that is, the desired style is to be changed to a technological feel, and calculate the corresponding confidence level; the intent category corresponding to "add texture" is "object addition" and calculate the corresponding confidence level, where object is used to indicate texture; the intent category corresponding to "adjust color brightness" is "attribute adjustment" and calculate the corresponding confidence level, where attribute is used to indicate color.
[0096] For example, the intent classifier can identify the intent category corresponding to "darken the background" as "attribute adjustment" and calculate the corresponding confidence score of 0.92, where the attribute is used to indicate brightness.
[0097] For example, the intent assigner can identify the intent category corresponding to "adjust the layout to twice the size" as "attribute adjustment" and calculate the corresponding confidence level as 0.88, where the attribute is used to indicate the layout.
[0098] S130. Combining the operation type and timestamp obtained in step S110, assign priority to the intent category obtained in step S120.
[0099] Alternatively, as an example, different priorities can be assigned to different operation types.
[0100] Alternatively, as another embodiment, based on the timestamp, the frequency corresponding to different operation objects under the same operation type is determined, and then higher priority (or higher weight) is assigned to high-frequency operation objects, and lower priority (or lower weight) is assigned to low-frequency operation objects.
[0101] For example, for the modification operation type, users "modify color" three times more frequently than "modify layout". Therefore, "modify color" can be assigned a high priority and "modify layout" can be assigned a low priority, so that color parameters can be quickly adjusted in subsequent generation.
[0102] If, after multiple rounds of dialogue, the frequency of "modify layout" is twice that of "modify color", then the operation of "modify layout" can be assigned a higher priority, and the operation of "modify color" can be assigned a lower priority.
[0103] Optionally, as another embodiment, based on the timestamp, the cumulative number of operations corresponding to different operation areas under the same operation type is determined. Then, operation areas with high cumulative operation counts are assigned higher priority (or higher weight), and operation areas with low cumulative operation counts are assigned lower priority (or lower weight).
[0104] For example, suppose a user adjusts the "poster title font" three times in a row. In this case, the operation type is modification, the operation area is the area where the "poster title font" is located, and the cumulative number of operations is 3, which is considered a high number. Therefore, a higher priority can be assigned to the area where the "poster title font" is located, and the font-related parameters can be preset in subsequent generation.
[0105] For example, if users click on hot zones frequently, higher priority can be assigned to the areas corresponding to the hot zones, and click operations in the hot zones can be pre-programmed to respond quickly in subsequent generation.
[0106] It should be understood that, in this application, the allocation of priority to intent categories is a dynamic process.
[0107] S140. The low-dimensional intent vector obtained in step S120 is encoded by combining image features and the assigned priority, and the generation instruction is output.
[0108] In this application, the compressed low-dimensional intent vector can be jointly encoded with image features (such as CLIP embedding) to generate context-aware generation instructions.
[0109] It should be understood that CLIP embedding refers to the process by which the CLIP (Contrastive Language–Image Pre-training) model converts images and text into vector representations. These vectors can reflect the semantic information of the images and text and reside in the same embedding space.
[0110] S150. Process according to the generation instructions and output the generation result.
[0111] For example, targeting Figure 4 For example, based on the generation instructions, a light blue logo can be generated, and matching fonts can be automatically recommended, such as a tech-style sans-serif font, reducing subsequent user operations.
[0112] In this embodiment of the application, by using the intent parsing module to compress multi-turn dialogues into low-dimensional intent vectors, the user's core intent and operational dependencies can be preserved.
[0113] Based on the above, Figure 6 This application provides another method for parsing dialogue intent.
[0114] Optionally, such as Figure 6 As shown, when the intent parsing module includes an intent conflict detection subunit, step S120 may include steps S121 to S126.
[0115] S121. The innovative algorithm subunit combines the input content and the history of multi-turn dialogues to calculate information entropy, and then filters high-value dialogue segments based on the information entropy.
[0116] S122, The intent conflict detection subunit identifies whether there is a contradiction between multiple rounds of high-value dialogue segments; if yes, then execute S123; if no, then execute S125.
[0117] S123. Initiate the follow-up questioning process and provide the user with multiple processing options.
[0118] S124. In response to the user's selection of a processing option, perform the processing indicated by the processing option for a multi-turn high-value dialogue segment.
[0119] For example, in the intent parsing module, the dialogue history compression unit extracts high-value dialogue fragments from the input content "generate a summer beach illustration" in the first round as "summer," "beach," and "illustration"; in the second round, it extracts high-value dialogue fragments from the input content "add snow and Christmas tree" as "snow" and "Christmas tree." For these high-value dialogue fragments, the intent conflict detection subunit identifies a semantic conflict between "summer" and "snow," and can initiate a follow-up questioning process, prompting the user to "detect a seasonal conflict, do you want to change to a winter scene or keep the summer theme?" and providing options for "winter scene," "keep summer theme," and "custom modification." When the user selects the "keep summer theme" option, the intent conflict detection subunit automatically ignores the high-value dialogue fragments "snow" and "Christmas tree" and provides replacement options again, such as "palm tree" and "umbrella."
[0120] Subsequently, based on the user's selection of replacement options, the corresponding segments of the selected replacement options can be added to high-value dialogue segments, and key operation nodes can be extracted and compressed together later.
[0121] S125. The compression algorithm subunit extracts key operation nodes from high-value dialogue segments, generates key operation chains, and compresses the key operation chains to generate low-dimensional intent vectors.
[0122] S126. The intent classifier identifies each operation node to determine its corresponding intent category and outputs the corresponding confidence score.
[0123] In this embodiment of the application, by introducing an intent conflict detection algorithm, contradictions between multiple rounds of instructions can be automatically identified and resolved.
[0124] Based on the dialogue intent parsing method described in the above embodiments, this application also provides an electronic device, including a processor and a memory, wherein the memory stores at least one computer instruction, which is loaded and executed by the processor to perform the steps performed by the server in the dialogue intent parsing method described in any of the above embodiments.
[0125] Based on the dialogue intent parsing method described in the above embodiments, this application also provides a computer-readable storage medium. For example, a non-transitory computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a CD-ROM, magnetic tape, a floppy disk, or an optical data storage device, etc. This storage medium stores computer instructions for executing the steps performed by the server in the dialogue intent parsing method described in any of the above embodiments, which will not be elaborated further here.
[0126] Based on the dialogue intent parsing method described in the above embodiments, this application also provides a computer-readable storage medium. For example, a non-transitory computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a CD-ROM, magnetic tape, a floppy disk, or an optical data storage device. This storage medium stores computer instructions for executing the steps performed by the zero terminal in the dialogue intent parsing method described in any of the above embodiments, which will not be elaborated further here.
[0127] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.
[0128] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.
Claims
1. A method for parsing dialogue intent, characterized in that, The method includes: Responding to input operations, retrieve the input content; By combining the input content and the content of multiple rounds of historical dialogue, a low-dimensional intent vector is parsed and obtained; The low-dimensional intent vector is encoded by combining image features, and a generation instruction is output. The generation instructions are processed to output the generation result; The multi-round historical dialogue content includes multi-round historical input content and multi-round historical generation results.
2. The method according to claim 1, characterized in that, After obtaining the input content, the process also includes: Record the operation type and timestamp; The operation types include at least: adding, modifying, and deleting.
3. The method according to claim 2, characterized in that, Combining the input content and multi-turn historical dialogue content, a low-dimensional intent vector is parsed and obtained, including: Calculate information entropy based on search input and multi-turn historical dialogue content; Based on the information entropy, high-value dialogue segments are selected; Extract key operation nodes from multi-round high-value dialogue segments to generate a key operation chain; The key operation chain is compressed into the low-dimensional intent vector.
4. The method according to claim 3, characterized in that, After parsing and passing the low-dimensional intent vector, the method further includes: For each key operational node, identify the corresponding intent category and confidence score; Priorities are assigned to intent categories based on the operation type and the timestamp. The step of encoding the low-dimensional intent vector in conjunction with image features and outputting generation instructions includes: The low-dimensional intent vector is encoded by combining the image features and the assigned priority, and the generation instruction is output.
5. The method according to claim 4, characterized in that, The intent categories include at least: object manipulation, attribute adjustment, and style transfer.
6. The method according to claim 3, characterized in that, After filtering high-value dialogue segments, the method further includes: Identify whether there are contradictions among the multi-turn high-value dialogue segments; If so, the follow-up questioning process will be initiated, and multiple processing options will be provided; In response to a selection operation for any processing option, the processing indicated by the processing option is performed on the multi-turn high-value dialogue segment.
7. An AI system, characterized in that, The system includes a multimodal input layer, an intent parsing module, a dynamic embedding engine, and a generation module connected in sequence. The multimodal input layer is used to respond to the input operation and obtain the input content; The intent parsing module is used to combine the input content and the content of multiple rounds of historical dialogue to parse and obtain a low-dimensional intent vector; The dynamic embedding engine is used to encode the low-dimensional intent vector in combination with image features and output generation instructions. The generation module is used to process the generation instructions and output the generation results; The multi-round historical dialogue content includes multi-round historical input content and multi-round historical generation results.
8. An electronic device, characterized in that, It includes a processor and a memory, wherein the memory stores at least one computer instruction, which is loaded and executed by the processor to perform the method of any one of claims 1 to 6.
9. A computer-readable storage medium, characterized in that, The storage medium stores at least one computer instruction, which is loaded and executed by a processor to perform the steps in the method of any one of claims 1 to 6.