Film and television script quality evaluation and optimization method for director creation process

Abstract

The application discloses a film and television script quality evaluation and optimization method for director creation process, which comprises the following steps: structurally segmenting input text, dividing the script into scenes, action units and shot candidate units; extracting cross-shot reserved fields from the shot candidate units, and setting source priority according to field source identification; constructing a field relationship graph in shot order, tracking the state change path of the same entity in the shot sequence; combining the action predicate to determine whether the state change is reasonable, and calculating the field missing risk, field conflict risk and field misediting risk; determining the minimum modifiable segment outside the text of the high-priority source field, generating a candidate modified text and reviewing the risk evaluation result to determine the output modification scheme. The method is beneficial to maintaining the visual continuity and director creation intention in the process of converting the script into a shot and a prompt text.

Description

Technical Field

[0001] This application relates to the field of film and television content production technology, specifically a method for evaluating and optimizing the quality of film and television scripts for the director's creative process. Background Technology

[0002] In the early stages of film and television production, the literary script is usually broken down into scenes, action descriptions, storyboard texts, and shot cues for use in the image or video generation process. With the application of automated storyboard and cue generation tools, the script text is repeatedly rewritten, supplemented, and compressed between different stages. Visual information such as characters, props, scenes, line of sight, spatial position, light source, and shot size is easily scattered across multiple text fragments.

[0003] Existing script processing methods often focus on text integrity, semantic fluency, or the generation of single-shot content, typically lacking tracking of changes in the state of the same entity between consecutive shots, and also lacking differentiation in the reliability of sources between the original script, director's intent, manually confirmed content, and automatically generated content. When automatically generated storyboards or prompts improperly supplement or replace information such as character positions, prop states, and action results, problems such as missing fields, inconsistencies, state jumps, or erroneous changes to director's intent can easily occur. These problems can lead to deviations in subsequent image or video generation results in terms of character relationships, prop transitions, spatial orientation, and action continuity, increasing the cost of manual review and repeated generation. Summary of the Invention

[0004] The purpose of this application is to provide a method for evaluating and optimizing the quality of film and television scripts in the director's creative process, so as to solve the problems mentioned in the background art.

[0005] According to one aspect of this application, a method for evaluating and optimizing the quality of film and television scripts in the director's creative process is provided, including the following steps: The input text is structured and segmented into scenes, action units, and shot candidate units. Extract cross-shot reserved fields from each shot candidate unit. Each cross-shot reserved field includes a field type, a field value, a shot number, and a field source identifier. Assign a source priority to each cross-shot reserved field based on the field source identifier. Organize all cross-shot reserved fields into a field relationship diagram according to the shot sequence to track the state change path of the same entity in the shot sequence; For each state change event, extract the action predicate from the action unit of the corresponding shot interval. If there is no action predicate to explain the state change, mark it as a state jump risk. Based on the field relationship diagram, source priority and state change path, calculate the field missing risk, field conflict risk and field erroneous modification risk. Based on source priority, the text containing high-priority source fields is marked as a modification protection zone. Outside the modification protection zone, the smallest modifiable fragment is determined and candidate modification text is generated. The candidate modified text is resubmitted for field extraction and the field relationship graph is updated. The risk assessment is re-executed, and the output modification plan is determined based on the risk assessment results.

[0006] Preferably, the text categories corresponding to the source priorities from high to low are: user original script, director's intention text, manually confirmed text, automatically generated storyboard, and automatically generated prompt words; when the user manually confirms the automatically generated storyboard or prompt words, the source priority of this field is raised to the level corresponding to the manually confirmed text, and the raised source priority participates in subsequent risk assessment and modification decisions.

[0007] Preferably, the field relationship graph is a heterogeneous graph, including shot nodes, character nodes, prop nodes, action nodes, scene nodes, and prompt word fragment nodes; the edge types between nodes include co-occurrence edges, adjacent connection edges, action-induced state change edges, character-prop association edges, and character-space direction association edges; each edge records its edge type, the identifiers of the two connected nodes, and the source priority of the fields involved; the adjacent connection edges connect nodes corresponding to the same entity between adjacent shots.

[0008] Preferably, a heterogeneous graph attention network is used to score the field relationship graph; the heterogeneous graph attention network includes an intra-type attention module, an inter-type attention module, and an edge scoring module connected in sequence; nodes of each type are mapped to feature vectors through their respective independent embedding layers and then input into the intra-type attention module, which calculates attention weights for nodes of the same type and outputs a weighted representation of each node; the inter-type attention module receives the output of the intra-type attention module and calculates the association representation between nodes of different types; the edge scoring module receives the output of the inter-type attention module and outputs a relationship quality score for each edge; based on the relationship quality score, the field missing risk, field conflict risk, and field erroneous modification risk are calculated.

[0009] Preferably, constructing the state change path includes: traversing pairs of the same entity nodes connected by adjacent receiving edges in the field relationship graph, reading the field values ​​of each node and arranging them in ascending order of shot number; if the field values ​​of the same entity are different in two adjacent shots, a state change event is recorded, and the state change event includes the state value before the change, the state value after the change, and the range of shot numbers involved.

[0010] Preferably, extracting the action predicate includes: performing dependency parsing on the text in the action unit to obtain predicate components, and using a language model to determine whether the predicate component has the semantic ability to change the state of a specified entity; determining whether there is an action predicate that explains the state change specifically means: if there is at least one action predicate in the action unit of the corresponding shot interval, and the semantic expression of the action predicate is the transition from the state before the change to the state after the change, then the state change is determined to be reasonable; otherwise, it is marked as a state jump risk.

[0011] Preferably, the risk of missing fields is calculated as follows: traverse the cross-shot reserved fields of the user's original script or director's intention text corresponding to the field source identifier, and check whether there is a matching record for the field in the storyboard candidate text or shot prompt. If not, it is marked as a risk of missing fields. The risk of conflicting fields is calculated as follows: compare the field values ​​of the same entity in adjacent shots. If the field values ​​are inconsistent and the corresponding state change event has been marked as a state jump risk, it is marked as a risk of conflicting fields. The risk of accidental modification of fields is calculated as follows: compare the field values ​​of the same entity in texts from different source categories. If the field value with lower source priority is inconsistent with the field value with higher source priority and there is no intermediate manual confirmation operation, it is marked as a risk of accidental modification of fields.

[0012] Preferably, determining the minimum modifiable segment includes: locating the candidate shot unit where the field with the risk marker is located, and using the candidate shot unit as a candidate modification area; when the source priorities of the conflicting parties are different, determining the text containing the field of the party with the lower source priority as the modification object; selecting a modification method based on the determination result of the state change path, wherein the modification method includes supplementing the action description at the state transition position, restoring the erroneously modified field to the original field value in the high-priority source, and deleting the automatically supplemented content that caused the conflict.

[0013] Preferably, when there are multiple candidate modified texts, after re-performing the risk assessment, the multiple candidate modified texts are sorted according to the following conditions: the candidate modified text with the fewest remaining risk markers after the re-examination is ranked first; when the number of remaining risk markers is the same, the candidate modified text with the smallest modification range is ranked first; when the modification range is the same, the candidate modified text that does not touch the modification protection zone is ranked first; wherein, the candidate modified text ranked first is the output modification scheme.

[0014] In another aspect, this application also provides a film and television script quality assessment and optimization system for the director's creative process, including: The structured segmentation module is used to segment the input text in a structured way, dividing the script into scenes, action units, and shot candidate units. The field extraction and priority setting module is used to extract cross-shot reserved fields from each shot candidate unit. Each cross-shot reserved field includes a field type, a field value, a shot number, and a field source identifier. The module sets a source priority for each cross-shot reserved field based on the field source identifier. The relationship building and path tracing module is used to organize all cross-shot reserved fields into a field relationship graph according to the shot sequence, and to trace the state change path of the same entity in the shot sequence. The risk assessment module is used to extract action predicates from the action units of the corresponding shot intervals for each state change event. If there is no action predicate to explain the state change, it is marked as a state jump risk. Based on the field relationship diagram, source priority and state change path, it calculates the field missing risk, field conflict risk and field erroneous modification risk. The candidate modification generation module is used to mark the text containing high-priority source fields as modification protection zones based on source priority, determine the smallest modifiable fragment outside the modification protection zones, and generate candidate modification text. The back-check output module is used to re-send the candidate modified text into the field extraction and update the field relationship graph, re-execute the risk assessment, and determine the output modification scheme based on the risk assessment results.

[0015] This application enables the structured recording and continuity verification of key visual information such as characters, props, scenes, action results, and spatial orientation during the conversion of film and television scripts into storyboard texts and shot prompts. This ensures that information scattered across different text sources has clear correspondences and continuity. By introducing source priority and modification protection zones, when the system detects missing, conflicting, or incorrectly modified fields, it can prioritize retaining the user's original script, director's intention text, and manually confirmed content, reducing deviations of automatically generated content from the creator's core expression. By using field relationship diagrams and state change paths to determine whether the changes of the same entity in adjacent shots have action basis, it can more accurately identify state jumps without reasonable continuity. Furthermore, candidate modified texts are re-extracted, relationship diagrams are updated, and risk assessments are conducted before determining the output scheme. This helps avoid introducing new inconsistencies through local corrections, thereby improving the stability of the script-to-storyboard and storyboard-to-shot prompt conversion and reducing the cost of repeated adjustments in subsequent image or video generation. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0017] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a schematic diagram illustrating a method for evaluating and optimizing the quality of film and television scripts in the director's creative process, as provided in an embodiment of this application.

[0018] Figure 2 This is a schematic diagram of the state change path tracking process provided in an embodiment of this disclosure.

[0019] Figure 3 This is a schematic diagram of the structure of a heterogeneous graph attention network provided in an embodiment of this disclosure.

[0020] Figure 4 This is a schematic diagram of the risk assessment process provided for an embodiment of this disclosure.

[0021] Figure 5 This is a schematic diagram of a film and television script quality assessment and optimization system for the director's creative process, provided as an embodiment of this application.

[0022] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0023] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0024] The film and television script quality assessment and optimization method provided in this disclosure, which is geared towards the director's creative process, is applicable to scenarios in film and television content production where script text is progressively converted into storyboard text, and then from storyboard text into image generation prompts or video generation prompts. The input to this method includes at least the user's original script, and may also include one or more of the following: director's intention text, candidate storyboard text, and shot prompts. The user's original script is written by the creator and records information such as scenes, characters, actions, and dialogue. The director's intention text contains the constraints given by the director regarding shot composition, character relationships, or mise-en-scène. The candidate storyboard text and shot prompts can be generated automatically or manually. After the above input text undergoes structured processing, field extraction, relationship modeling, risk assessment, and local correction by this method, a modified scheme that has passed visual continuity verification is output. This modified scheme is used in subsequent image or video generation stages.

[0025] The following detailed description, in conjunction with specific embodiments, illustrates the implementation process of the film and television script quality assessment and optimization method for the director's creative process described in this application. It should be noted that these embodiments are merely for explaining this application and not for limiting its scope of protection. Any conventional adjustments or substitutions made by those skilled in the art to the steps without departing from the concept of this application should be included within the scope of protection of this application.

[0026] like Figure 1 As shown in the illustration, this application discloses a schematic diagram of a method for evaluating and optimizing the quality of film and television scripts in the director's creative process, including the following method steps: S1, perform structured segmentation on the input text, dividing the script into scenes, action units, and shot candidate units; S2, extract cross-shot reserved fields from each shot candidate unit. Each cross-shot reserved field includes a field type, a field value, the shot number it belongs to, and a field source identifier; set a source priority for each cross-shot reserved field according to the field source identifier; S3 organizes all cross-shot reserved fields into a field relationship diagram according to the shot sequence, and tracks the state change path of the same entity in the shot sequence; S4. For each state change event, extract the action predicate from the action unit of the corresponding shot interval. If there is no action predicate to explain the state change, mark it as a state jump risk. Based on the field relationship diagram, source priority and state change path, calculate the field missing risk, field conflict risk and field erroneous modification risk. S5, mark the text containing the high-priority source field as the modification protection zone according to the source priority, determine the smallest modifiable segment outside the modification protection zone and generate candidate modified text; S6, the candidate modified text is resubmitted to the field extraction and the field relationship graph is updated, the risk assessment is re-executed, and the output modification scheme is determined based on the risk assessment results.

[0027] In some embodiments, for step S1, the input text is structurally segmented, dividing the script into three levels: scenes, action units, and shot candidate units. A scene is a segment in the script divided according to time or space transitions, with each scene corresponding to a continuously occurring scene fragment. An action unit is the smallest action description within a scene, containing four elements: the executing subject, the action, the object, and the visual result. For example, in the sentence "She put the letter in her pocket," the executing subject is "she," the action is "put it in," the object is "letter," and the visual result is "the letter is in her pocket." Shot candidate units correspond to the smallest text fragments that may subsequently generate images or videos, with each shot candidate unit containing at least one action unit or one static scene description. The segmentation process utilizes the structured output capability of a language model: the system sends the script text to the language model, and the language model returns scene boundary markers, the extraction results of the four elements of each action unit, and the segmentation results of shot candidate units. The segmentation granularity can be adjusted according to the script type and shot density.

[0028] In some embodiments, for step S2, after completing the structured segmentation, the system extracts cross-shot reserved fields from each shot candidate unit. Cross-shot reserved fields refer to visual information that should not change arbitrarily in a group of consecutive shots, or is only allowed to change when a reasonable action occurs. The reason for introducing this concept is that information such as character position, prop holding status, and gaze direction in the script is scattered in various paragraphs of text. If they are not extracted into structured records one by one, subsequent comparisons can only be made at the level of the entire text, making it difficult to accurately locate which specific visual information has become abnormal between shots.

[0029] The types of fields retained across shots cover characters, props, scenes, actions, character positions, gaze directions, prop states, light sources, time periods, shot sizes, camera positions, and camera movements. Among them, the character field records the identifier of the character appearing in the shot; the prop field records key items and their current state; the scene field marks the spatial environment in which the shot is located; the action field records the action content that occurs in the current shot; the character position field indicates the relative position of the character in the frame space; the gaze direction field indicates the direction the character is looking; the prop state field describes how the prop is held, its placement, or its open / closed state; the light source field marks the direction of the light source in the frame; and the shot size, camera position, and camera movement fields record the framing, shooting angle, and camera movement method, respectively. Each extracted cross-shot retained field contains at least four attributes: field type, field value, shot number, and field source identifier. The field source identifier is used to record in which type of input text the field information first appeared, and this identifier is used as a classification basis when setting the source priority in subsequent processes.

[0030] The extraction process calls the structured output interface of the language model; the system sends the text of the shot candidate unit along with the field type definition and output format requirements to the language model, and the language model returns the type, value and source label of each field in the unit; if multiple characters or props appear in the same shot, the system generates independent field records for each character and each prop.

[0031] In one embodiment, the system sets source priority for each cross-shot retention field based on the field source identifier. The reason for setting source priority is that the reliability of information carried by different text sources varies objectively; the actions and states explicitly written by the creator in the original script usually have higher certainty; while the storyboard descriptions or prompts supplemented by the automatic generation module may deviate from the creator's original intention. If the source of information is not distinguished in the subsequent conflict determination, the system will not be able to determine which side should be retained when faced with two contradictory records.

[0032] For example, there are five source categories, ranked from highest to lowest priority: user-generated original script, director's intention text, manual confirmation text, automatically generated storyboard, and automatically generated prompts. Among them, actions and states explicitly written in the user-generated original script have the highest priority, the constraints given in the director's intention text have the second highest priority, and the content supplemented by the generation module in the automatically generated storyboard and automatically generated prompts has a lower priority. Manual confirmation text corresponds to a special case: when a user manually confirms or edits a storyboard or prompt after the automatically generated result is produced, the source priority of that field is upgraded from the original automatic generation level to the manual confirmation text level, and the upgraded source priority participates in all subsequent risk assessments and modification decisions.

[0033] In some embodiments, the system can quantify source priority into a credibility value. The credibility value is determined according to the order of source priority. For example, the user's original script, director's intention text, manual confirmation text, automatically generated storyboard, and automatically generated prompt words can be mapped to 5, 4, 3, 2, and 1, respectively. Alternatively, it can be normalized to a value between 0 and 1 without changing the above order. The initial credibility value can be preset according to project requirements or statistically adjusted during use based on the historical frequency of correction of this type of source. The quantified credibility value is written into the record of the cross-shot retention field for use in the edge scoring calculation of the subsequent field relationship graph.

[0034] In some embodiments, for step S3, such as Figure 2 As shown, Figure 2 This is a schematic diagram of the state change path tracking process provided in the embodiments of this disclosure. In S201, all cross-shot reserved fields are organized into a field relationship diagram according to the shot order.

[0035] After completing field extraction and source priority settings, the system organizes all cross-shot retained fields into a field relationship graph according to the shot order. The field relationship graph adopts a heterogeneous graph structure, which contains multiple types of nodes and multiple types of edges.

[0036] Specifically, there are six types of nodes: Shot nodes correspond to each shot candidate unit, recording the shot number and scene; Character nodes correspond to each character instance appearing in the shot; Prop nodes correspond to each key prop instance; Action nodes correspond to the action records extracted from the action unit; Scene nodes correspond to the spatial environment in which the shot is located; and Cue word fragment nodes correspond to the generated shot cue word text.

[0037] There are five edge types: co-occurrence edges connect nodes that appear together in the same shot; adjacent connection edges connect nodes corresponding to the same entity between adjacent shots, such as the line connecting the "envelope" prop node in shot 3 and the "envelope" prop node in shot 4; action-induced state change edges connect action nodes to prop nodes or character position nodes whose state changes due to the action; character-prop association edges indicate the relationship of a character holding, using, or placing a prop; character-spatial direction association edges indicate the correspondence between the character's position and orientation and the scene space; each edge records its edge type, the identifiers of the two nodes it connects, and the source priority of the fields involved, with the source priority being the lower value of the fields at both ends of the connection.

[0038] In S202, a heterogeneous graph attention network is used to score the field relationship graph and output the relationship quality score of each edge. The lower the score, the higher the risk of visual continuity between the two fields connected by the edge.

[0039] Among them, such as Figure 3 As shown, Figure 3 This is a schematic diagram of the heterogeneous graph attention network provided in this embodiment of the disclosure. Specifically, it includes an embedding layer, an intra-type attention module, an inter-type attention module, and an edge scoring module. Specifically, the embedding layer sets up an independent embedding layer for each of the six node types. Each embedding layer maps the field value text representation of the node of that type to a feature vector of fixed dimension. The embedding dimension can be configured to be 128-dimensional or 256-dimensional. The input of the embedding layer is the word index sequence obtained after word segmentation of the node's field value text, and the output is a node feature vector of fixed dimension.

[0040] The intra-type attention module receives all node feature vectors output from the embedding layer, groups them by node type, and performs multi-head self-attention computation for nodes of the same type within the same shot or between adjacent shots: it measures the degree of association between nodes of the same type using a dot product attention mechanism of query vector, key vector, and value vector, and outputs the weighted representation of each node in the group of nodes of the same type. In some embodiments, the number of attention heads can be configured to 4 or 8. The role of intra-type attention is to allow nodes of the same type to refer to each other, for example, to allow multiple character nodes in the same scene to adjust their respective representation weights according to the range of shots in which they appear.

[0041] The inter-type attention module receives the output of the intra-type attention module and calculates the association strength between nodes of different types. This module uses edge type as an intermediary for information propagation: for each edge in the field relationship graph, it reads the representations of the nodes at both ends of the edge after intra-type attention processing, calculates the cross-type attention weights through a linear transformation and Softmax normalization, and then aggregates the information of neighboring nodes by weighted summation, outputting a node representation that integrates cross-type associations. After processing by this module, the representation of each node contains information of both nodes of the same type and nodes of different types.

[0042] The edge scoring module receives the node representations output by the inter-type attention module. For each edge in the field relationship graph, this module concatenates the representation vectors of the nodes at both ends of the edge and inputs them into a fully connected layer. Then, it is mapped to a value between 0 and 1 by the Sigmoid activation function and outputs the relationship quality score of the edge. The meaning of the relationship quality score is: close to 1 indicates that the state transmission between the field pairs is reasonable, and close to 0 indicates that there is a high risk of visual continuity. This score is used in the calculation of field missing risk, field conflict risk, and field erroneous modification risk at each stage.

[0043] Optionally, regarding the training process, the training data for this network comes from self-annotated film and television scripts and storyboard datasets. The annotations include whether there are field conflicts, missing fields, or incorrect field modifications between adjacent shots, with annotation values ​​of 0 or 1. During training, the annotation values ​​are converted into relation quality labels that are consistent with the meaning of relation quality scores, where 1 indicates that there are no field conflicts, missing fields, or incorrect field modifications between the corresponding fields of the edge, and 0 indicates that there are at least one of the following: field conflicts, missing fields, or incorrect field modifications between the corresponding fields of the edge.

[0044] During training, the relation quality label of each edge is used as a supervision signal. The difference between the predicted score and the relation quality label is calculated using the binary cross-entropy loss function and backpropagated to update the network parameters. The Adam optimizer is used for training, and the learning rate can be configured to be on the order of 0.001. The training epochs are stopped when the loss value on the validation set no longer decreases.

[0045] In S203, the system tracks the state change path of the same entity in the sequence of shots. After the field relationship graph is established, the system tracks the state change path of the same entity in the sequence of shots. The state change path describes the state evolution process of the same character, the same prop, or the same spatial relationship from one shot to the next. Taking a letter as an example, its state may be "on the table", "in the hand", and "in the pocket", corresponding to the 1st, 2nd, and 3rd shots, respectively.

[0046] The construction process is as follows: The system traverses all pairs of the same entity nodes connected by adjacent receiving edges in the field relationship graph, reads the field values ​​of each node and arranges them in ascending order of shot number. If the field values ​​of the same entity are different in two adjacent shots, the system records a state change event. Each state change event contains three pieces of information: the state value before the change, the state value after the change, and the range of shot numbers involved. After arranging all the state change events of the same entity in order of shot number, the complete state change path of the entity is formed.

[0047] In some embodiments, for step S4, such as Figure 4 As shown, Figure 4 This is a schematic diagram of the risk assessment process provided in an embodiment of the present disclosure. In S401, for each state change event, the action predicate is extracted from the action unit of the corresponding shot interval.

[0048] For each state change event, the system needs to determine whether there is an action unit within the involved shot range that can explain the change. The basis for this determination is the action predicate extracted from the action unit.

[0049] Specifically, the extraction of action predicates consists of two steps. First, dependency parsing is performed on the text in the action unit to identify the predicate components in the sentence. Dependency parsing is a fundamental technology in natural language processing. Its principle is to establish a dominance or subordination relationship between each word in the sentence and other words, thereby identifying the predicate headword and its argument structure. Second, the language model determines whether the predicate component has the semantic ability to change the state of the specified entity.

[0050] Specifically, the system submits the state value before the change, the state value after the change, and the candidate predicate to the language model. The language model determines whether the predicate can semantically express the transition from the previous state to the next state. For example, the predicate "stuff" can express the transition from "in hand" to "in pocket", while the predicate "watch" cannot express the change in the position of the prop. In some embodiments, an action predicate classifier can also be trained using existing film and television storyboard data to perform this judgment. The input of the classifier is the predicate text and entity state pair, and the output is a binary label indicating whether they match.

[0051] The specific judgment rule is as follows: if there is at least one action predicate in the action unit of the corresponding shot interval, and the semantics of the action predicate can express the transition from the state before the change to the state after the change, the system determines that the state change is reasonable; if there is no such action predicate, the system marks the state change event as a state jump risk.

[0052] To illustrate how this judgment works, consider a specific scenario: the second shot shows "the man holding a key," and the third shot's cue changes to "the key is in the lock." The system searches for predicates in the action units corresponding to the second and third shots. If it finds the action unit "he inserts the key into the lock," the predicate "inserts" semantically explains the change in the key's position from "in his hand" to "on the lock," and the system deems it reasonable. If no action unit containing such a predicate exists within the interval, the system marks "key position change" as a state transition risk. Compared to simply checking whether two texts contain the same keywords, this judgment method based on action sequence provides a finer granularity for identifying visual consistency issues in consecutive shots.

[0053] In S402, based on the field relationship diagram, source priority, and state change path, the risks of field missing, field conflict, and field erroneous modification are calculated.

[0054] Specifically, to address the risk of missing fields, the system iterates through all cross-shot reserved fields corresponding to the user's original script or director's intention text with the field source identifier, and checks whether there is a matching record for the field in the storyboard candidate text or shot cue words. The matching is determined by field type and entity identifier. For fields such as light source, time period, shot type, camera position, and camera movement that do not have independent entity identifiers, the matching is determined by field type, shot number, scene, and corresponding field value.

[0055] If a character, prop, scene, action, character position, gaze direction, prop status, light source, time period, shot type, camera position, or camera movement appears in the script but has no corresponding record in subsequent storyboards or prompts, the system marks it as a field missing risk. The risk level is related to the source priority of the field, as well as whether the field belongs to the core action object in the current scene. The core action object is the entity corresponding to the "object" in the four elements of the action unit, or the entity that is directly affected by the action predicate and produces a visual result. The risk level is higher when the core action object is missing.

[0056] For field conflict risks, the system traverses all adjacent connecting edges in the field relationship graph and compares the field values ​​of the nodes at both ends of the edge. If the field values ​​at both ends are inconsistent, and if the corresponding state change event has been marked as a state transition risk, then the field pair is marked as a field conflict risk, and the conflict risk level of the field pair is increased according to the state transition risk. If the change has been determined to be a reasonable change supported by an action predicate, then no conflict risk is marked. Optionally, as boundary protection, the system sets a few additional rules to handle obviously mutually exclusive cases. For example, the same character in the same shot cannot be in two mutually exclusive spatial positions at the same time, and the same prop cannot be in both "open" and "closed" states at the same time.

[0057] Regarding the risk of accidental modification of fields, the system compares the field values ​​of the same entity in texts from different source categories. If the field value with lower source priority (such as records in automatically generated storyboards or automatically generated prompts) is inconsistent with the field value with higher source priority (such as records in user original scripts or director intent texts), and there is no record of user manual confirmation between the two records, the system marks it as a risk of accidental modification of fields. The level of accidental modification risk depends on the difference in source priority between the two ends of the modified field: the greater the difference, the higher the risk level.

[0058] In one embodiment, the level of the above three types of risks takes into account the priority of the field source, the position of the field in the shot sequence, whether there is a reasonable state change path, and the degree of correlation between the field and the core action of the current scene.

[0059] The preset threshold for the relationship quality score is determined between 0 and 1, with a default value of 0.5. In embodiments with a validation set, this preset threshold can also be determined based on the comprehensive accuracy of the identification results for field missing risk, field conflict risk, and field erroneous modification risk in the validation set, and recorded in the system configuration. Field pairs with relationship quality scores below the preset threshold do not generate new risk types independently, but if a field missing risk, field conflict risk, or field erroneous modification risk marker has already been formed, the corresponding risk level will be increased by one level; if the field pair has not formed the above risk marker, it will only be recorded as a low relationship quality field pair.

[0060] In one embodiment, a special case exists during risk assessment: the script does not explicitly state the subsequent state of an entity. For example, the script might say "she turned and left," but not specify whether the letter is still in her hand. In this case, the system does not directly add the unstated state as a confirmed fact, nor does it directly classify it as a conflict. If the entity has a clear field value in the previous shot and is mentioned again in the next shot, but its state information is not explicitly given by any input text, the system marks the field as a confirmed void. A confirmed void means that the current state of the field cannot be uniquely determined from existing text and needs to be handled with minimal changes during the optimization phase. When generating candidate modified text, the system prioritizes adding a statement consistent with the entity's field value in the previous shot to confirmed voids. For example, if the previous shot records "she held the letter in her hand," but the next shot does not mention the letter's state, the system prioritizes adding "she still held the letter and turned away" rather than automatically assigning a new position to the letter.

[0061] In some embodiments, for step S5, before generating the modification scheme, the system first marks the modification protection zone. The modification protection zone is a text processing range restriction, which is used to constrain the operable range when generating candidate modifications. The system reads all field records of the source priority corresponding to the user's original script and director's intention text, and marks the text span of these fields and their directly associated action units as modification protection zones. If the user subsequently manually confirms a piece of automatically generated content, the text span of that content is also included in the modification protection zone. The marking granularity is based on the text span of the field. When the system cannot determine the specific text span of a field in the input text, it will designate the shot candidate unit where the field is located as the modification protection zone. Low-priority source field text in a shot candidate unit that does not fall into the modification protection zone can still be used as a candidate modification object without changing the content of the modification protection zone.

[0062] After modifying the protected area marker, the system determines the smallest modifiable fragment outside the protected area. The smallest modifiable fragment is the shortest text area that can fix missing fields, field conflicts, or field errors, without changing the plot order, director's intentions, or modifying the text content within the protected area.

[0063] Specifically, the process begins by locating the candidate shot unit containing the risk-marked field and designating this unit as the candidate modification area. When the source priorities of the conflicting parties differ, the text containing the field from the lower-priority source is identified as the modification target. When the source priorities are the same, the party to be modified is determined based on its consistency with the state change path of the preceding and following shots. Finally, the modification method is selected based on the determination result of the state change path. These modification methods include: if a missing action sequence causes a state jump, adding an action description near the jump location; if there is an erroneous modification, restoring the erroneously modified field to its original value from the higher-priority source; and if there is a conflict and the priorities of both parties are similar, deleting the automatically supplemented content causing the conflict or replacing it with a statement consistent with the preceding text.

[0064] In some embodiments, for step S6, the system, based on the constraints of the minimum modifiable segment and the modification protection zone, calls the language model to generate candidate modified text; wherein, the input received by the language model includes: the candidate shot unit or cue word segment to be modified, the risk type marked on the segment, the modification protection zone marker, and the field records of the preceding and following shots; the language model generates one or more candidate modification schemes under the constraint instructions, each scheme only makes local adjustments within the minimum modifiable segment; the constraint instructions require that the language model not change any content within the modification protection zone, not use the text within the modification protection zone as a replacement, deletion or addition object, not change the plot order, not add characters or props not mentioned in the script, and the modified text should be consistent with the field records of the preceding and following shots in terms of entity state, spatial direction and action result.

[0065] In one embodiment, the system performs a backtesting verification on each candidate modified text. The backtesting process involves re-extracting fields from the modified text, re-extracting cross-shot retained fields, and setting source priorities for the modified content. For newly generated or replaced fields in the candidate modified text, the field source identifier is recorded according to the source of the storyboard candidate text or shot cue word to which the candidate modified text belongs. If the field value is recovered from a high-priority source field, its corresponding high-priority source field identifier is also recorded as a comparison basis, but this does not elevate the candidate modified text itself to a high-priority source. Based on this, the field relationship diagram is updated, and three types of risk assessments are re-executed. The purpose of the backtesting is to confirm whether the original risk markers have been eliminated and whether the modification has introduced new field conflicts or missing fields.

[0066] The specific rules for re-inspection are as follows: if a candidate modification eliminates the original risk marker and does not generate a new risk marker, the candidate modification passes the re-inspection; if it eliminates the original risk but introduces a new risk, such as fixing the conflict in the position of the letter but causing the character's line of sight to be inconsistent with the previous shot, the system will lower the ranking of the candidate modification.

[0067] In one embodiment, when multiple candidate modified texts exist, the system sorts them hierarchically according to the following conditions: the one with the fewest remaining risk markers after re-checking is ranked first; if the remaining number is the same, the one with the smallest modification magnitude is ranked first; the modification magnitude is compared in the following order: the number of changed characters in the candidate modified text relative to the original text to be modified, the number of cross-shot reserved fields that have changed, and the number of shot candidate units affected; if the first item is the same, the second item is compared, and the candidate modified text whose three items are smaller or the same is considered to have a smaller or the same modification magnitude; if the modification magnitude is the same, the one that has not touched the modification protection zone is ranked first.

[0068] Among them, touching the modification protection zone means that the candidate modified text uses the text within the modification protection zone as an object to be deleted, replaced or added, or causes the field type, field value, shot number or field source identifier of the field within the modification protection zone to change after a review; using the text within the modification protection zone only as context input without changing its text content and field extraction results does not fall under touching the modification protection zone; among them, the candidate modified text ranked first is the output modification scheme.

[0069] In some embodiments, the language model adopts a pre-trained language model based on the Transformer architecture, such as the Tongyi Qianwen Big Model, etc. This disclosure does not limit the specific model type, and fine-tunes the instructions for the above tasks; the language model adopts a decoder-only Transformer architecture, which consists of four parts connected in sequence: a token embedding layer, a position encoding module, a multi-layer stacked Transformer decoding block, and an output projection layer.

[0070] The token embedding layer receives the token index sequence obtained after the input text is processed by the word segmenter, and maps each token index to an embedding vector of fixed dimension. The embedding dimension can be configured to 4096 dimensions. The word segmenter uses the byte pair encoding method to build a vocabulary, and the vocabulary size can be configured to be in the tens of thousands.

[0071] The positional encoding module uses Rotated Positional Encoding (RoPE). The principle of RoPE is to group the query vector and key vector by dimension and apply a position-related rotation transformation to each group, so that the attention weights of the two tokens naturally contain information about their relative positions. This encoding method does not require additional trainable parameters and directly affects the query vector and key vector for attention calculation in subsequent decoding blocks.

[0072] The Transformer decoding block is the main part of the model, consisting of multiple stacked layers, configurable to 32 or more. Each decoding block contains three sub-modules, connected in the following order: a masked multi-head self-attention sub-module, a feedforward network (FFN) sub-module, and two layer normalization points. The specific connection relationships are as follows: The masked multi-head self-attention submodule receives the input after the first layer normalization and generates a query matrix, a key matrix, and a value matrix through three sets of linear projections. The number of attention heads can be configured to 32. In each attention head, the query matrix and key matrix are first subjected to rotational position encoding via RoPE, then the scaled dot product attention is calculated, and a causal mask is applied to ensure that each position only focuses on the current position and the tokens before it. The outputs of each attention head are concatenated and restored to the original dimension through a linear projection layer. The output of this submodule is added to the input of the submodule through a residual connection.

[0073] The feedforward network submodule receives the result of the residual output of the masked multi-head self-attention submodule after the second layer normalization. This submodule adopts the SwiGLU activation structure, which specifically includes three linear projections: the input vector is passed through two parallel linear projections, the output of one path is passed through the SiLU activation function (i.e., the Swish activation function), and then multiplied element-wise with the output of the other path. The result of the multiplication is restored to the original dimension by the third linear projection. The intermediate dimension of the feedforward network is usually configured to be several times the embedding dimension, for example, it can be configured to 11008 dimensions. The output of the feedforward network submodule is also added to its input through the residual connection, which is used as the final output of the decoding block and passed to the next decoding block.

[0074] After all the decoding blocks are processed, the output of the last decoding block is normalized once and then enters the output projection layer. The output projection layer is a linear projection that maps the hidden state to the vocabulary dimension. The probability distribution of the next token is then obtained through the Softmax function. During inference, the model generates tokens one by one in an autoregressive manner until the output terminator is reached or the maximum generation length is reached.

[0075] In one embodiment, the pre-training phase is completed using a large-scale general text corpus. This phase is a well-known language model pre-training process in the industry. The key to this embodiment is the instruction fine-tuning phase, which constructs training samples for four types of tasks and performs multi-task joint fine-tuning on the same model.

[0076] The training samples for the four types of tasks involved in this embodiment are composed as follows: The first category is text structured segmentation tasks. The input of the training samples is a piece of film and television script text with task instructions. The model is required to output four elements: scene boundary markers, execution subject, action, object and visual result of each action unit, as well as the segmentation results of shot candidate units. The output format is JSON structure, and the annotation data comes from manual annotation of the scenes and action units of existing film and television scripts.

[0077] The second category is the cross-shot preserved field extraction task. The input of the training samples is the text of a single shot candidate unit, along with a list of field type definitions and task instructions. The model is required to output the type, value, and source annotation of each cross-shot preserved field in the unit. The output is also in JSON structure, and the annotation data comes from the field-by-field annotation of the visual information in the shot candidate unit by human.

[0078] The third category is the action predicate semantic judgment task. The input of the training sample contains three parts: a candidate predicate, a state value before the change and a state value after the change of a state change event. The task instruction requires the model to judge whether the predicate can semantically express the transition from the state before the change to the state after the change. The output is a binary judgment result. The labeled data comes from the manual annotation of the matching relationship between common action predicates and entity state change pairs in film and television scenes.

[0079] The fourth category is candidate modified text generation tasks. The input of the training samples includes the candidate unit text of the shot to be modified, risk type markers, modified protected area labels, and field records of adjacent shots before and after. The task instructions require the model to make local adjustments to the smallest modifiable segment under constraints and output the modified text. The labeled data comes from manual modification demonstrations of existing conflict cases.

[0080] During fine-tuning, the four types of training samples are mixed and trained uniformly. The training objective is the standard autoregressive language modeling objective, which is to maximize the conditional log-likelihood of each token in the output part of the training samples. The loss function is calculated only for the tokens in the output part, and the input and instruction parts are not included in the loss calculation. The Adam optimizer is used for fine-tuning, and the learning rate can be configured to be on the order of 1e-5 to 2e-5. A cosine annealing scheduling strategy is adopted. The number of training rounds stops when the overall accuracy of each task on the validation set no longer increases. During the fine-tuning process, the LoRA method can be used to reduce the amount of parameter updates: only the low-rank adaptation matrix is ​​inserted on the query projection and value projection of the attention sub-module of each decoding block, and the other parameters are frozen and not updated. The rank of LoRA can be configured to 16 or 32.

[0081] After fine-tuning, the model automatically generates results in the corresponding task output format based on the task instruction identifier carried in the input text during the inference phase. For tasks that require structured output (the first three types), the system performs JSON parsing after receiving the model output. For candidate modified text generation tasks, the system directly reads the text content output by the model.

[0082] Using the above methods, the system shifts the focus of script quality assessment from the fluency of the entire text to the stability of the transmission of specific visual fields in continuous shots. When determining conflicts, it determines the direction of correction based on the source priority. During modification generation, it preserves the creator's core intent by modifying the protection zone and the constraints of the minimum modifiable segment. This method can reduce the probability of inconsistencies in characters, props, spatial orientation, and action states during the generation of images or videos in continuous shots, reduce the number of repeated generations caused by contradictions between cue words, and improve the stability of the automated conversion from script to storyboard and from storyboard to cue words.

[0083] It should be noted that although the operations of the method of this application are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. On the contrary, the steps depicted in the flowchart can be performed in a different order. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0084] Please see Figure 5 , Figure 5 This application provides a film and television script quality assessment and optimization system for the director's creative process. The system embodiments are related to... Figure 1 Corresponding to the illustrated method embodiments, this system can be specifically applied to various electronic devices. The system specifically includes: The structured segmentation module 501 is used to perform structured segmentation on the input text, dividing the script into scenes, action units, and shot candidate units. The field extraction and priority setting module 502 is used to extract cross-shot reserved fields from each shot candidate unit. Each cross-shot reserved field includes a field type, a field value, a shot number, and a field source identifier. The module sets a source priority for each cross-shot reserved field according to the field source identifier. The relation building and path tracing module 503 is used to organize all cross-shot reserved fields into a field relationship graph according to the shot sequence, and to trace the state change path of the same entity in the shot sequence. The risk assessment module 504 is used to extract action predicates from the action units of the corresponding shot intervals for each state change event. If there is no action predicate to explain the state change, it is marked as a state jump risk. Based on the field relationship diagram, source priority and state change path, it calculates the field missing risk, field conflict risk and field erroneous modification risk. The candidate modification generation module 505 is used to mark the text containing the high-priority source field as the modification protection zone according to the source priority, determine the smallest modifiable segment outside the modification protection zone and generate candidate modification text; The back-check output module 506 is used to re-send the candidate modified text into the field extraction and update the field relationship diagram, re-execute the risk assessment, and determine the output modification scheme based on the risk assessment results.

[0085] Each processing unit and / or module in the embodiments of this application can be implemented by an analog circuit that implements the functions described in the embodiments of this application, or by software that executes the functions described in the embodiments of this application.

[0086] Based on the same inventive concept, this application also provides an electronic device. The method corresponding to the electronic device can be the method in the foregoing embodiments, and its problem-solving principle is similar to that method. The electronic device provided in this application includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the methods and / or technical solutions of the foregoing embodiments of this application.

[0087] Figure 6 The diagram illustrates the structure of an electronic device suitable for implementing the methods and / or technical solutions in the embodiments of this application. The electronic device includes a central processing unit (CPU) 601, which can perform various appropriate actions and processes based on a program stored in a read-only memory (ROM) 602 or a program loaded from a storage section 608 into a random access memory (RAM) 603. The RAM 603 also stores various programs and data required for system operation. The CPU 601, ROM 602, and RAM 603 are interconnected via a bus 604. An input section 606, an output section 607, a communication section 609, and an input / output (I / O) interface 605 are also connected to the bus 604.

[0088] In particular, the methods and / or embodiments in this application can be implemented as computer software programs. For example, the embodiments disclosed in this application include a computer program product comprising a computer program carried on a storage medium, the computer program containing program code for performing the methods shown in the flowchart. When the computer program is executed by the central processing unit (CPU) 601, it performs the functions defined in the methods of this application.

[0089] Another embodiment of this application provides a computer-readable storage medium having computer program instructions stored thereon, which can be executed by a processor to implement the methods and / or technical solutions of any one or more embodiments of this application described above.

[0090] The flowcharts or block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of electronic devices, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-specific system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0091] The above description represents the preferred embodiments of the present invention. It should be noted that, for those skilled in the art, various improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for evaluating and optimizing the quality of film and television scripts in the director's creative process, characterized in that, include: The input text is structured and segmented into scenes, action units, and shot candidate units. Extract cross-shot reserved fields from each shot candidate unit. Each cross-shot reserved field includes a field type, a field value, a shot number, and a field source identifier. Assign a source priority to each cross-shot reserved field based on the field source identifier. Organize all cross-shot reserved fields into a field relationship diagram according to the shot sequence to track the state change path of the same entity in the shot sequence; For each state change event, extract the action predicate from the action unit of the corresponding shot interval. If there is no action predicate to explain the state change, mark it as a state jump risk. Based on the field relationship diagram, source priority and state change path, calculate the field missing risk, field conflict risk and field erroneous modification risk. Based on source priority, the text containing high-priority source fields is marked as a modification protection zone. Outside the modification protection zone, the smallest modifiable fragment is determined and candidate modification text is generated. The candidate modified text is resubmitted for field extraction and the field relationship graph is updated. The risk assessment is re-executed, and the output modification plan is determined based on the risk assessment results.

2. The method for evaluating and optimizing film and television script quality based on the director's creative process as described in claim 1, characterized in that, The source priority, from highest to lowest, corresponds to the following text categories: user-generated original script, director's intention text, manually confirmed text, automatically generated storyboard, and automatically generated prompts. When a user manually confirms an automatically generated storyboard or prompt, the source priority of that field is raised to the level corresponding to the manually confirmed text. The raised source priority then participates in subsequent risk assessments and modification decisions.

3. The method for evaluating and optimizing film and television script quality based on the director's creative process as described in claim 1, characterized in that, The field relationship graph is a heterogeneous graph, containing shot nodes, character nodes, prop nodes, action nodes, scene nodes, and prompt word fragment nodes; the edge types between nodes include co-occurrence edges, adjacent connection edges, action-induced state change edges, character-prop association edges, and character-space direction association edges; each edge records its edge type, the identifiers of the two connected nodes, and the source priority of the fields involved; the adjacent connection edges connect nodes corresponding to the same entity between adjacent shots.

4. The method for evaluating and optimizing film and television script quality based on the director's creative process as described in claim 3, characterized in that, A heterogeneous graph attention network is used to score the field relationship graph. The heterogeneous graph attention network includes an intra-type attention module, an inter-type attention module, and an edge scoring module connected in sequence. Nodes of each type are mapped to feature vectors through their respective independent embedding layers and then input into the intra-type attention module. The intra-type attention module calculates attention weights for nodes of the same type and outputs a weighted representation of each node. The inter-type attention module receives the output of the intra-type attention module and calculates the association representation between nodes of different types. The edge scoring module receives the output of the inter-type attention module and outputs a relationship quality score for each edge. The risk of missing fields, conflicting fields, and accidental modification of fields are calculated based on the relationship quality score.

5. The method for evaluating and optimizing the quality of film and television scripts based on the director's creative process as described in claim 3, characterized in that, Constructing the state change path includes: traversing pairs of the same entity nodes connected by adjacent receiving edges in the field relationship graph, reading the field values ​​of each node and arranging them in ascending order of shot number; if the field values ​​of the same entity are different in two adjacent shots, a state change event is recorded, and the state change event includes the state value before the change, the state value after the change, and the range of shot numbers involved.

6. The method for evaluating and optimizing the quality of film and television scripts based on the director's creative process as described in claim 1, characterized in that, Extracting the action predicate includes: performing dependency parsing on the text in the action unit to obtain the predicate component, and using a language model to determine whether the predicate component has the semantic ability to change the state of a specified entity; determining whether there is an action predicate that explains the state change specifically means: if there is at least one action predicate in the action unit of the corresponding shot interval, and the semantic expression of the action predicate is the transition from the state before the change to the state after the change, then the state change is determined to be reasonable; otherwise, it is marked as a state jump risk.

7. The method for evaluating and optimizing film and television script quality based on the director's creative process according to claim 2, characterized in that, The risk of missing fields is calculated as follows: traverse the cross-shot reserved fields of the user's original script or director's intention text corresponding to the field source identifier, and check whether there is a matching record for the field in the storyboard candidate text or shot prompt. If not, it is marked as a field missing risk. The risk of conflicting fields is calculated as follows: compare the field values ​​of the same entity in adjacent shots. If the field values ​​are inconsistent and the corresponding state change event has been marked as a state jump risk, it is marked as a field conflict risk. The risk of accidental modification of fields is calculated as follows: compare the field values ​​of the same entity in texts from different source categories. If the field value with lower source priority is inconsistent with the field value with higher source priority and there is no intermediate manual confirmation operation, it is marked as a field accidental modification risk.

8. The method for evaluating and optimizing the quality of film and television scripts based on the director's creative process according to claim 1, characterized in that, Determining the minimum modifiable segment includes: locating the candidate shot unit where the field with the risk marker is located, and using the candidate shot unit as the candidate modification area; when the source priorities of the conflicting parties are different, determining the text containing the field of the party with the lower source priority as the modification object; selecting the modification method based on the determination result of the state change path, wherein the modification method includes supplementing the action description at the state transition position, restoring the erroneously modified field to the original field value in the high-priority source, and deleting the automatically supplemented content that caused the conflict.

9. The method for evaluating and optimizing the quality of film and television scripts based on the director's creative process as described in claim 1, characterized in that, When multiple candidate modified texts exist, after re-performing the risk assessment, the candidate modified texts are sorted according to the following conditions: the candidate modified text with the fewest remaining risk markers after the re-examination is ranked first; when the number of remaining risk markers is the same, the candidate modified text with the smallest modification range is ranked first; when the modification range is the same, the candidate modified text that does not touch the modification protection zone is ranked first; among them, the candidate modified text ranked first is the output modification scheme.

10. A film and television script quality evaluation and optimization system for the director's creative process, characterized in that, include: The structured segmentation module is used to segment the input text in a structured way, dividing the script into scenes, action units, and shot candidate units. The field extraction and priority setting module is used to extract cross-shot reserved fields from each shot candidate unit. Each cross-shot reserved field includes a field type, a field value, a shot number, and a field source identifier. The module sets a source priority for each cross-shot reserved field based on the field source identifier. The relationship building and path tracing module is used to organize all cross-shot reserved fields into a field relationship graph according to the shot sequence, and to trace the state change path of the same entity in the shot sequence. The risk assessment module is used to extract action predicates from the action units of the corresponding shot intervals for each state change event. If there is no action predicate to explain the state change, it is marked as a state jump risk. Based on the field relationship diagram, source priority and state change path, it calculates the field missing risk, field conflict risk and field erroneous modification risk. The candidate modification generation module is used to mark the text containing high-priority source fields as modification protection zones based on source priority, determine the smallest modifiable fragment outside the modification protection zones, and generate candidate modification text. The back-check output module is used to re-send the candidate modified text into the field extraction and update the field relationship graph, re-execute the risk assessment, and determine the output modification scheme based on the risk assessment results.

Images