Commodity video intelligent generation method based on multi-modal analysis and dynamic narrative architecture

Abstract

The invention relates to the technical field of video generation, in particular to an intelligent commodity video generation method based on multi-modal analysis and a dynamic narrative architecture, which comprises the following steps: inputting original commodity data and a user behavior log, and outputting a dynamic selling point weight vector through a selling point value evaluation engine; s2, inputting the dynamic selling point weight vector output in S1 into a narrative flow generator, and executing the following steps: intercepting first K core selling points according to the weight vector to form a narrative trunk chain; based on the historical preference data of the user, inserting an emotion enhancement node to generate a branch enhancement narrative flow; generating a multi-version narrative flow instruction set in combination with the real-time network bandwidth data; retrieving a preset video clip library according to the instruction set to generate an initial video sequence; detecting semantic fault regions between adjacent segments to generate compensation animation parameters; and injecting compensation animation parameters and outputting a continuous narrative video stream. According to the invention, the jamming feeling and the frame skipping phenomenon caused by the change of a narrative structure or different sources of video clips are reduced, and the overall watching experience of a user is improved.

Description

Technical Field

[0001] The present invention relates to the field of video generation technology, and in particular to a method for intelligently generating commodity videos based on multimodal analysis and dynamic narrative architecture. Background Art

[0002] With the rapid development of e-commerce and short video platforms, product videos have become an important way to showcase products online. However, existing product video generation technologies generally have limitations.

[0003] Most product videos use a templated structure, showcasing product features through a fixed sequence of shots. This lacks personalized adjustments tailored to individual user interests, resulting in poor alignment between video content and user interests and inefficient information delivery. Traditional product videos often construct narratives based on static scripts or a pre-set sequence of selling points, failing to incorporate real-time user behavior data or the complexity of product attributes for flexible editing. This results in some users' core selling points being overlooked, impacting video retention and conversion rates.

[0004] Existing video stitching methods often suffer from discontinuities in content style, emotional tone, or tempo when assembling multiple product videos. This can lead to visual jumps, abrupt rhythmic changes, and other discontinuities during viewing. Some technologies attempt to mitigate this problem using simple transition templates, but these lack transition strategies based on semantic judgment and motion trajectory adaptation, resulting in limited compensation. Summary of the Invention

[0005] The present invention provides an intelligent product video generation method based on multimodal analysis and dynamic narrative architecture, integrating product image and text semantics, user behavior preferences and network resource conditions to achieve a dynamically adjustable, clearly structured, emotionally linked, and continuous and natural video generation method.

[0006] The intelligent product video generation method based on multimodal analysis and dynamic narrative architecture includes the following steps:

[0007] S1, dynamic selling point value assessment of products: Input the original product data and user behavior logs, and output the dynamic selling point weight vector through the selling point value assessment engine, including:

[0008] Parse product image and text descriptions to generate basic function tag sets;

[0009] Integrate user stay time and click heat map to generate attention distribution matrix;

[0010] Calculate the real-time value weight of each selling point based on the behavior log and generate a dynamic selling point weight vector;

[0011] S2, attention-driven narrative flow generation: Input the dynamic selling point weight vector output by S1 into the narrative flow generator and execute:

[0012] S21, intercept the first K core selling points according to the weight vector to form the narrative backbone chain;

[0013] S22, inserting emotion reinforcement nodes based on user historical preference data to generate branched enhanced narrative flow;

[0014] S23, generates a multi-version narrative flow instruction set based on real-time network bandwidth data;

[0015] S3, fault-compensated video synthesis: the narrative stream instruction set generated by S2 is input into the video synthesizer, and the preset video clip library is retrieved according to the instruction set to generate the initial video sequence; the semantic fault areas between adjacent clips are detected to generate compensating animation parameters; the compensating animation parameters are injected to output a continuous narrative video stream.

[0016] Optionally, the step of parsing the product image and text description to generate a basic function tag set in S1 specifically includes:

[0017] Parse the graphic and text descriptions of the products, including structured parameter information and natural language description text, and use named entity recognition technology to extract key entity information in the products, including materials, functions, and applicable scenarios, and mark their corresponding semantic categories. Perform dependency syntax analysis on the text to identify the semantic dependency relationship between functional vocabulary and its modified components, and then construct a function-attribute association tree, establish semantic binding for each function point and its specific attributes, and organize the extraction results into a standardized set of basic function labels through semantic normalization, synonym merging and context consistency verification.

[0018] Optionally, the construction of the attention distribution matrix in S1 includes:

[0019] Based on user behavior log data, a two-dimensional matrix reflecting the distribution of user attention is constructed. The dwell time series and click heat map data are extracted from the interactive behavior of users when browsing product pages. The page display area is divided into multiple grid units, and the click frequency and gaze dwell time in each grid unit are counted. By fusing the two sources of attention, click behavior and time behavior, the attention density value of each grid unit is calculated to form a space-time two-dimensional attention distribution matrix, which is used to characterize the degree of user attention in different areas of the product.

[0020] Optionally, the calculation of the real-time value weight in S1 specifically includes:

[0021] Obtain the basic function tag set and attention distribution matrix, and associate the tag information with the user's attention behavior; through the front-end embedding mechanism and interface element recognition, obtain the display area position of each basic function tag on the page, and establish a mapping relationship between tags and areas. In the area corresponding to each tag, count the user's click behavior and attention distribution, and comprehensively calculate the real-time value weight of the tag to measure the user's current attention to the selling point. It also serves as a driving factor for narrative flow generation, and finally outputs a dynamic weight vector including all basic function tags.

[0022] Optionally, the construction of the S21 narrative backbone chain includes: receiving a dynamic selling point weight vector, sorting each selling point according to the value weight, calculating a truncation threshold for truncation based on the attribute complexity of the product, extracting the first several key selling points to construct a backbone narrative chain, and forming a linear narrative backbone chain of the core functions of the product.

[0023] Optionally, S22 includes the insertion of emotion reinforcement nodes, specifically including generating an emotion preference vector based on the user's historical behavior data, judging whether each narrative node in the narrative trunk chain meets the set emotion activation conditions, and if so, inserting an emotion reinforcement node that matches the user's preference style at the corresponding position, expanding a branch-enhanced narrative flow including a branch path.

[0024] Optionally, the generation of the multi-version narrative flow instruction set in S23 includes: matching the corresponding visual special effects level standard according to the real-time network bandwidth status of the current terminal device, traversing all narrative nodes in the branch-enhanced narrative flow, the branch-enhanced narrative flow includes the main chain node and the branch nodes inserted by the emotion enhancement mechanism, allocating corresponding display time and lens special effects level to each narrative node, generating structured rendering instructions, and finally outputting multiple multi-version narrative flow instruction sets adapted to different bandwidth conditions, generating a triplet for each narrative node: [node ID, display time, special effects level].

[0025] Optionally, the generation of the initial video sequence in S3 includes inputting a multi-version narrative flow instruction set into a video synthesizer, parsing the node identifier, display duration and special effects level information included in each instruction, retrieving the corresponding video unit from a preset video clip library according to the node identifier, and assembling them in accordance with the instruction sequence to construct an initial video sequence including multiple video units, wherein the video units carry their start and end times, visual special effects tags and semantic description information.

[0026] Optionally, the S3 also includes semantic fault detection and compensation, specifically including performing fault detection on adjacent video units in the initial video sequence, comprehensively considering the two dimensions of semantic similarity and temporal continuity, and calculating the coherence score between adjacent units; when the coherence score is lower than the set score threshold, it is determined that there is a fault area, and compensation animation parameters are generated accordingly, including compensation type, compensation duration and transition trajectory obtained based on optical flow estimation between adjacent frames.

[0027] Optionally, the S3 includes inserting the compensating animation between the fault areas, constructing a transition frame sequence, and recombining it with the original video unit to generate a complete video stream; at the same time, adjusting the duration and aligning the beats of the audio track to output a continuous narrative video stream with consistent rhythm and visual continuity.

[0028] Beneficial effects of the present invention:

[0029] This invention, through a multimodal evaluation system comprised of image and text analysis, user behavior fusion, and tag region mapping, dynamically generates a selling point weight vector that reflects the current user's attention preferences. This weight vector responds in real time to data such as user dwell time and click behavior, ensuring that narrative content unfolds in the order that best aligns with user interests. Compared to traditional static ranking methods, this dynamic weighting mechanism significantly increases the display priority of key selling points, achieving a narrative-driven conversion rate improvement.

[0030] This invention incorporates user emotional preference vectors into product video generation, constructing a branching narrative flow with emotion-reinforcing nodes. When significant emotional preferences are met, the system automatically matches stylized nodes that align with the user's aesthetic preferences and inserts them into the main narrative chain, creating a dual-track presentation structure. This structure effectively addresses the existing issues of lack of personalized information and monotonous content expression, enhancing user resonance and immersion without interfering with the primary message.

[0031] This paper addresses the issue of content jumps in video narratives by proposing a dual-factor fault detection model based on semantic similarity and temporal continuity. It also introduces a dynamic compensation animation generation mechanism. The system adaptively selects the compensation type (scaling, displacement, or gradient) based on the scoring results, and uses optical flow estimation to generate natural transition trajectories, achieving smooth transitions between fault regions. Furthermore, a sound and video synchronization mechanism ensures visual and rhythmic consistency in the transitions. This significantly reduces the perceived stuttering and frame skipping caused by changes in narrative structure or different video clip sources, improving the overall viewing experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0033] Figure 1 Schematic diagram of a method flow in an embodiment of the present invention. DETAILED DESCRIPTION

[0034] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. Those skilled in the art may also implement some known technologies in other alternative ways. The accompanying drawings are only for describing the embodiments in more detail and are not intended to limit the present invention in any specific way.

[0035] like Figure 1 As shown in FIG, the intelligent product video generation method based on multimodal analysis and dynamic narrative architecture includes the following steps:

[0036] S1, dynamic selling point value assessment of products: Input the original product data and user behavior logs, and output the dynamic selling point weight vector through the selling point value assessment engine, including:

[0037] Parse product image and text descriptions to generate basic function tag sets;

[0038] Integrate user stay time and click heat map to generate attention distribution matrix;

[0039] Calculate the real-time value weight of each selling point based on the behavior log and generate a dynamic selling point weight vector. The specific solution is as follows S11-S13:

[0040] S11, basic function tag set generation: parse the structured data and unstructured text in the product description and perform the following processing:

[0041] Named Entity Recognition (NER) technology is used to extract entity information such as product materials, functions, and applicable scenarios. Based on dependency syntax analysis, a semantic dependency structure between functions and attributes is constructed to form a function-attribute association tree, and a standardized basic function label set is output: ;in, Indicates the functional tag set of the product. For the Function tags.

[0042] S12, attention distribution matrix construction: Fusion of user real-time behavior logs generates a spatial-temporal two-dimensional attention matrix. The steps are as follows:

[0043] Extract the sequence of user's stay time on the current product page: , Indicates the total length of the sequence of user's stay time on the current product page (i.e. the number of stay time segments);

[0044] Capture the coordinate data of user clicks on the heat map: , Indicates the total number of click events (i.e., the number of click points) that occurred on the current product page;

[0045] Divide the product display interface into The grid, each grid cell is , calculate its attention density value:

[0046] ;

[0047] in, are the row and column indices of the grid respectively, Indicates the index of the p-th click point in the click heat map, ranging from 1≤p≤k. Represents the index of the qth dwell time segment during which the sight stays in the grid cell Inside, Represents the total index of all dwell time segments, ranging from 1≤r≤m. is the indicator function, if the point Falling on the grid If yes, the value is 1, otherwise it is 0. Indicates that it falls on the grid The index set of stay time periods in , is an adjustable attention fusion weight factor, which is used to balance the contribution of click behavior and dwell time, and finally obtain the attention distribution matrix: .

[0048] S13, real-time value weight calculation: based on basic function tag set and the attention distribution matrix , calculate the real-time value weight of each tag:

[0049] Establish a mapping relationship between labels and screen areas: ;in, Representation Label The display area in the interface.

[0050] Calculate the weight value for each label:

[0051] ;

[0052] in, Represents the coordinates of a single click point Number of clicks, is the total number of click points, represents the total attention intensity within the label region, Represents the adjustment factor that controls the weight distribution between click behavior and attention hotspots, Respectively represent the row and column traversal index of the grid matrix (used to traverse the global grid). The final output is the real-time value weight:

[0053] , each Indicates function label The real-time value weight of It is a dynamic weight vector calculated based on user attention behavior and product display area.

[0054] In addition, regarding the extraction of core product information based on named entity recognition, the specific operations are as follows:

[0055] When processing product descriptions, the system feeds the text portion into a pre-trained named entity recognition model to identify key information related to the core features of the product, including:

[0056] Material entities: such as "pure cotton", "stainless steel", "ABS plastic", etc., representing the composition of the product;

[0057] Functional entities: such as "heat preservation", "antibacterial", "shock absorption", "sun protection", etc., which represent the functional use of the product;

[0058] Applicable scenario entities: such as "outdoor cycling", "baby products", "business meetings", etc., reflecting the usage environment or group of people.

[0059] To achieve the above functions, the NER model adopts a structure based on BERT or BiLSTM-CRF, combined with a product category segmentation vocabulary for training and adaptation, so that it can accurately extract target entities in multiple types of product copy. When outputting, the model labels each recognized entity with its category (such as [material], [function], [scene]).

[0060] After completing NER extraction, we further perform dependency parsing on the product description text. The main goal is to explore the semantic structure of "which specific attributes correspond to a certain function." The specific operations for constructing the "function-attribute" semantic dependency structure based on dependency parsing are as follows:

[0061] Use a dependency parser (SpaCy, HanLP, or Stanza) to construct a grammatical tree for the product description sentence and extract syntactic dependencies (such as subject, predicate, modifier, complement, etc.). Using the identified "functional words" as anchors, expand their dependency modification chains outward to capture the associated attribute words or modifiers, for example:

[0062] Input: "This water bottle is made of double-walled stainless steel and has good heat preservation effect."

[0063] Functional anchor: "insulation"

[0064] Dependency path analysis: "insulation" ← "effect" ← "have", forming the semantic unit "insulation effect"

[0065] Attribute association: "Double-layer stainless steel structure" and "insulation" establish a structural connection through syntactic dependency

[0066] The above semantic relationship is abstracted into a "function-attribute association tree": the node is the function keyword, and the child node is its related attributes and modified content.

[0067] After obtaining the "function-attribute association tree", perform the following standardization process:

[0068] Function trunk extraction: For each function semantic tree, extract its root node as a candidate function label;

[0069] Semantic normalization: Use word vector clustering or synonym dictionaries (e.g., "antibacterial," "antibacterial," and "sterilization" are grouped as the same function) for classification and merging.

[0070] Attribute completion and fusion: While retaining the core semantics of the function tag, it combines its modified attributes to generate a composite tag (such as "insulation_double-layer stainless steel");

[0071] Multi-source consistency check: If the product description also contains structured parameters (such as specifications and function parameter tables), it is used to supplement or correct the generated label;

[0072] Output format: The final output is a set of standardized, machine-readable functional labels.

[0073] Create a tag With screen area The mapping relationship between them is to locate each function tag displayed on the product page to its specific display area on the user interface, so as to calculate the attention and weight of the tag by combining the user click heat map and dwell behavior. The specific implementation process is as follows:

[0074] 1. Interface element positioning: On the front-end product detail page, HTML structures are used to display product features, such as keywords in the graphic description area, function columns in the parameter form, and selling point modules represented by visual icons. These elements often have unique DOM node IDs or CSS class names, and their display area on the screen can be captured using front-end monitoring code (such as JavaScript tracking scripts).

[0075] 2. Display area coordinate extraction: When the product page is loaded, call the getBoundingClientRect() interface provided by the browser to obtain the location information of the element corresponding to each function label, including its upper left corner coordinates, width and height, that is:

[0076] Horizontal start and end (xmin, xmax);

[0077] Vertical start and end (ymin, ymax);

[0078] This rectangle defines the label Display area on the screen .

[0079] 3. Matching label content with the function label set: After executing S11, a standardized function label set 𝑇 is obtained. At this time, the text labels displayed on the front-end page need to be compared with this set. If the interface text is consistent with a label in the label set (or a successful match is achieved through synonym merging), a mapping relationship is established;

[0080] The successfully matched tag ti is bound to the rectangular area Ri where it is located.

[0081] 4. Construct mapping function: finally form the mapping function from label to region , that is, the system can find the display area of ​​any function label in the interface, and use it to subsequently calculate the user's behavioral indicators such as click density and dwell time in the area.

[0082] S2, attention-driven narrative flow generation: Input the dynamic selling point weight vector output by S1 into the narrative flow generator and execute:

[0083] S21, intercept the first K core selling points according to the weight vector to form the narrative backbone chain;

[0084] S22, inserting emotion reinforcement nodes based on user historical preference data to generate branched enhanced narrative flow;

[0085] S23, generates a multi-version narrative flow instruction set in combination with real-time network bandwidth data.

[0086] The S2 specifically includes the following steps:

[0087] S21, the narrative backbone chain is constructed based on the dynamic selling point weight vector: ; Sort by weight from high to low to generate a selling point priority queue:

[0088] ;

[0089] Dynamically calculate the backbone chain truncation threshold based on the complexity of product attributes:

[0090] ; Among them, attribute complexity = number of function labels User cognitive difficulty coefficient, Indicates the ceiling function. Before extraction The tags with the highest weights constitute the main narrative chain:

[0091] .

[0092] S22, emotional enhancement node insertion: Based on user historical preference data, enhance the expressiveness of the main chain:

[0093] First, analyze the user's historical behavior data and generate the sentiment preference vector: , Indicates the user's preference for technology-themed content, with a value range of [0, 1]. Indicates the user's preference for luxury content, with a value range of [0, 1]. Indicates the user's preference for minimalist content, with a value range of [0,1];

[0094] For each node in the backbone chain , Indicates the first Function tags to determine the emotional activation conditions:

[0095] If satisfied: ;

[0096] Then insert the emotion reinforcement node at the corresponding position: ;

[0097] At the same time, the corresponding branch path is generated: ;

[0098] The final output is a branch-enhanced narrative flow with emotional branches : Branching Path .

[0099] Indicates the maximum value among the three style preference values, represents the style category (technology / luxury / simple) that maximizes the emotional preference value. Indicates the inserted emotion reinforcement node, is the emotion activation threshold, which is used to determine whether to insert an emotion node, and its value is 0.7.

[0100] Emotional activation threshold It is necessary to ensure the emotional trigger threshold of "preference significance". The purpose of the emotional activation mechanism is to insert the corresponding emotional reinforcement node only when the user has a clear tendency towards a certain type of emotion. If the threshold is too low, it is easy to cause users without obvious preferences to frequently trigger branches, interfere with the expression of the main narrative chain, and destroy the concentration of information transmission.

[0101] The maximum value in the normalized vector ≥ 0.7 is regarded as the judgment criterion of "preference dominance".

[0102] Experimental verification: User behavior data analysis based on A / B testing

[0103] In the product prototype test, we conducted A / B group experiments on multiple emotion activation thresholds (0.5, 0.6, 0.7, and 0.8) to obtain the following threshold-emotion trigger comparison data:

[0104] Table 1 Threshold-Emotional Trigger Comparison Table

[0105] Threshold θ Emotional branch trigger rate User retention rate Branch bounce rate (noise) 0.5 High (>60%) middle High (>35%) 0.6 middle middle middle 0.7 Moderate (≈35%) High (↑11%) Low (<15%) 0.8 Low (<20%) middle Low

[0106] The results show that setting θ = 0.7 can moderately exert the personalized narrative effect without excessively interfering with the main chain, and is the optimal balance between user acceptance and information completeness.

[0107] S23, multi-version instruction set compilation: based on real-time network bandwidth data ,Compile multiple versions of special effect instruction sets for each node in the narrative flow, and define the mapping relationship between special effect level and bandwidth:

[0108] Table 2 Mapping relationship between special effect level and bandwidth

[0109] Bandwidth range (Mbps) Special Effects Level Lens type B≥20 L3 3D particle effects 10≤B<20 L2 2D dynamic graphics B<10 L1 Static image and text translation

[0110] Traversing branches enhances narrative flow ,Generate a triple for each narrative node: [node ID, display duration, special effect level];

[0111] Output a complete multi-version narrative flow instruction set: ; Each Corresponding special effects level The video rendering instruction set below.

[0112] S3, fault-compensated video synthesis: The narrative flow instruction set generated in S2 is input into the video synthesizer, and the preset video clip library is retrieved according to the instruction set to generate the initial video sequence; the semantic fault areas between adjacent clips are detected to generate compensation animation parameters; the compensation animation parameters are injected to output a continuous narrative video stream. Specifically, it includes S31-S33:

[0113] S31, Initial Video Sequence Generation: Transforming the Narrative Flow Instruction Set Enter the Video Compositor and do the following:

[0114] Parse each instruction triplet: [node ID, display duration, special effect level]; retrieve the corresponding atomic video unit from the video clip library based on the node ID; assemble according to the instruction sequence to generate the initial video sequence: ; Each Carry the following metadata items:

[0115] [Start and end timestamps, special effects markers ; Indicates the first Atomic video units, For video units The semantic descriptor vector of .

[0116] S32, semantic fault detection and compensation: Perform fault detection on adjacent units in the initial video sequence, and the operation is as follows:

[0117] For each pair of adjacent cells ,Calculate the semantic coherence score:

[0118] If the rating , then it is determined that there is a fault, and it is necessary to insert a compensation animation and generate compensation animation parameters. Represents a video unit The end timestamp of Represents a video unit The starting timestamp of is the maximum time interval allowed, used to normalize the timing difference, is the weighted coefficient of semantic similarity and temporal consistency in coherence scoring, For the Score the semantic coherence of adjacent video units, It is the fault judgment threshold of the coherence score, and its value ranges from 0.6 to 0.7.

[0119] S33, continuous video stream reconstruction: repair the fault area according to the compensation animation parameters and output the final video: insert a transition frame sequence at each fault position ,Build the reconstruction video:

[0120] ; Among them, the symbol Represents the splicing operation of video clips; synchronizes the time length of the audio track and animation compensation, and aligns the sound and picture as a whole; outputs the final continuous narrative video stream, Represents a sequence of compensated animation frames, inserted in between, To inject the complete continuous narrative video stream after compensation, Represents a narrative flow instruction set The total number of nodes in , including the sum of the number of atomic video units corresponding to the main chain nodes and the emotional branch nodes.

[0121] Select the compensation animation type:

[0122] Based on fault scoring The value range of , calls a compensation type selector to automatically determine which type of compensation animation to use. The judgment rules are as follows:

[0123] If the score is below a first threshold of 0.4 ( ), select zoom compensation, which is suitable for situations where semantic jumps are drastic and screen transitions are abrupt;

[0124] If the score is in the middle range ( 6): Use displacement compensation, which is suitable for slightly misplaced narrative jumps;

[0125] If the score is close to but not at the threshold ( ): Gradual compensation is selected for smooth transitions with slight semantic changes.

[0126] Dynamically calculate the compensation animation duration. The duration of the transition animation is automatically calculated based on the fault score. The lower the score, the more serious the fault, and the corresponding compensation animation duration is longer. The "inverse proportional enhancement" method is adopted, that is, the lower the score, the longer the compensation time, to ensure that the picture transition is natural enough.

[0127] Finally, the pixel motion trajectory between the two key frames of the previous and next video units is obtained through optical flow estimation. The specific method is as follows:

[0128] extract The last frame of The dense optical flow algorithm is used to calculate the motion vector field between the two frames of images. The main direction, speed and structural changes are extracted from the motion vector as the trajectory parameters of the compensation animation. This trajectory is used to guide interpolation or special effects rendering to achieve a seamless transition between the two video contents.

[0129] The present invention encompasses any alternatives, modifications, equivalents, and solutions that fall within the spirit and scope of the present invention. To provide a thorough understanding of the present invention, specific details are described in detail below in connection with the preferred embodiments of the present invention, but those skilled in the art will be able to fully understand the present invention without these detailed descriptions. Furthermore, to avoid unnecessary confusion regarding the essence of the present invention, well-known methods, processes, procedures, components, and circuits have not been described in detail.

[0130] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. An intelligent product video generation method based on multimodal analysis and dynamic narrative architecture, characterized by: The following steps are involved: S1, dynamic selling point value assessment of products: Input the original product data and user behavior logs, and output the dynamic selling point weight vector through the selling point value assessment engine, including: Parse product image and text descriptions to generate basic function tag sets; Integrate user stay time and click heat map to generate attention distribution matrix; Calculate the real-time value weight of each selling point based on the behavior log and generate a dynamic selling point weight vector; S2, attention-driven narrative flow generation: Input the dynamic selling point weight vector output by S1 into the narrative flow generator and execute: S21, intercept the first K core selling points according to the weight vector to form the narrative backbone chain; S22, inserting emotion reinforcement nodes based on user historical preference data to generate branched enhanced narrative flow; S23, generates a multi-version narrative flow instruction set based on real-time network bandwidth data; S3, fault-compensated video synthesis: the narrative stream instruction set generated by S2 is input into the video synthesizer, and the preset video clip library is retrieved according to the instruction set to generate the initial video sequence; the semantic fault areas between adjacent clips are detected to generate compensating animation parameters; the compensating animation parameters are injected to output a continuous narrative video stream.

2. The method for intelligently generating product videos based on multimodal analysis and dynamic narrative architecture according to claim 1, characterized in that: The generation of a basic function tag set by parsing the product image and text description in S1 specifically includes: Parse the graphic and text descriptions of the products, including structured parameter information and natural language description text, and use named entity recognition technology to extract key entity information in the products, including materials, functions, and applicable scenarios, and mark their corresponding semantic categories. Perform dependency syntax analysis on the text to identify the semantic dependency relationship between functional vocabulary and its modified components, and then construct a function-attribute association tree, establish semantic binding for each function point and its specific attributes, and organize the extraction results into a standardized set of basic function labels through semantic normalization, synonym merging and context consistency verification.

3. The method for intelligently generating product videos based on multimodal analysis and dynamic narrative architecture according to claim 2, characterized in that: The construction of the attention distribution matrix in S1 includes: Based on user behavior log data, a two-dimensional matrix reflecting the distribution of user attention is constructed. The dwell time series and click heat map data are extracted from the interactive behavior of users when browsing product pages. The page display area is divided into multiple grid units, and the click frequency and gaze dwell time in each grid unit are counted. By fusing the two sources of attention, click behavior and time behavior, the attention density value of each grid unit is calculated to form a space-time two-dimensional attention distribution matrix, which is used to characterize the degree of user attention in different areas of the product.

4. The method for intelligently generating product videos based on multimodal analysis and dynamic narrative architecture according to claim 2, characterized in that: The calculation of the real-time value weight in S1 specifically includes: Obtain the basic function tag set and attention distribution matrix, and associate the tag information with the user's attention behavior; through the front-end embedding mechanism and interface element recognition, obtain the display area position of each basic function tag on the page, and establish a mapping relationship between tags and areas. In the area corresponding to each tag, count the user's click behavior and attention distribution, and comprehensively calculate the real-time value weight of the tag to measure the user's current attention to the selling point. It also serves as a driving factor for narrative flow generation, and finally outputs a dynamic weight vector including all basic function tags.

5. The method for intelligently generating product videos based on multimodal analysis and dynamic narrative architecture according to claim 1, characterized in that: The construction of the S21 narrative backbone chain includes: receiving a dynamic selling point weight vector, sorting each selling point according to the value weight, calculating a truncation threshold for truncation based on the attribute complexity of the product, extracting the first several key selling points to construct a backbone narrative chain, and forming a linear narrative backbone chain of the core functions of the product.

6. The method for intelligently generating product videos based on multimodal analysis and dynamic narrative architecture according to claim 5, characterized in that: The S22 includes the insertion of emotion reinforcement nodes, specifically including generating an emotion preference vector based on the user's historical behavior data, judging whether each narrative node in the narrative trunk chain meets the set emotion activation conditions, and if so, inserting an emotion reinforcement node that matches the user's preference style at the corresponding position, expanding a branch-enhanced narrative flow including a branch path.

7. The method for intelligently generating product videos based on multimodal analysis and dynamic narrative architecture according to claim 6, characterized in that: The generation of the multi-version narrative flow instruction set in S23 includes: matching the corresponding visual special effects level standard according to the real-time network bandwidth status of the current terminal device, traversing all narrative nodes in the branch-enhanced narrative flow, the branch-enhanced narrative flow includes the main chain node and the branch nodes inserted by the emotion enhancement mechanism, allocating corresponding display time and lens special effects level to each narrative node, generating structured rendering instructions, and finally outputting multiple multi-version narrative flow instruction sets adapted to different bandwidth conditions, and generating a triplet for each narrative node: [node ID, display time, special effects level].

8. The method for intelligently generating product videos based on multimodal analysis and dynamic narrative architecture according to claim 7, characterized in that: The generation of the initial video sequence in S3 includes inputting a multi-version narrative flow instruction set into a video synthesizer, parsing the node identifier, display duration and special effect level information included in each instruction, retrieving the corresponding video unit from a preset video clip library according to the node identifier, and assembling them in accordance with the instruction sequence to construct an initial video sequence including multiple video units, wherein the video units carry their start and end times, visual special effects tags and semantic description information.

9. The method for intelligently generating product videos based on multimodal analysis and dynamic narrative architecture according to claim 8, characterized in that: The S3 also includes semantic fault detection and compensation, which specifically includes performing fault detection on adjacent video units in the initial video sequence, comprehensively considering the two dimensions of semantic similarity and temporal continuity, and calculating the coherence score between adjacent units; when the coherence score is lower than the set score threshold, it is determined that there is a fault area, and compensation animation parameters are generated accordingly, including compensation type, compensation duration, and transition trajectory obtained based on optical flow estimation between adjacent frames.

10. The method for intelligently generating product videos based on multimodal analysis and dynamic narrative architecture according to claim 8, characterized in that: The S3 includes inserting the compensating animation between the fault areas, constructing a transition frame sequence, and recombining it with the original video unit to generate a complete video stream; at the same time, adjusting the duration and aligning the beats of the audio track to output a continuous narrative video stream with consistent rhythm and visual continuity.

Images