Information processing method and device, equipment and medium product

Abstract

The invention discloses an information processing method and device, equipment and a medium product, and the method comprises the steps: obtaining a plurality of pieces of progress record information of a sample media item, calling a language model to carry out summarization based on the plurality of pieces of progress record information of the sample media item, and obtaining an initial item progress summary; based on the time information and the execution progress described by each clause in the initial project progress summary, calibrating a plurality of clauses in the initial project progress summary to obtain a target project progress summary, a plurality of clauses in the target project progress summary are arranged according to a time sequence, and a logical relationship among execution progresses described by the plurality of clauses is matched with a logical relationship specified by a media production process, and performing model training on a language model by adopting the target project progress summary, the trained language model can generate the project progress summary of the corresponding media project based on the multiple pieces of progress record information of any media project, so that the generation efficiency and accuracy of the project progress summary are improved, and the labor cost is saved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of Internet, in particular to the technical field of computer, and more particularly to an information processing method and device, equipment and medium product. BACKGROUND

[0002] At present, the process of generating a project progress summary of a media project is roughly as follows: first, the progress information of the media project is manually collected by each party of manpower of the media project; then, the progress information collected by each party of manpower is followed up and managed in real time by professional personnel, and each progress information is uniformly sorted and summarized, so as to artificially generate a project progress summary based on the result of sorting and summarizing. It can be seen that the process of generating a project progress summary currently needs to rely on the artificial participation of professional personnel, which not only leads to time and labor consumption due to low artificial efficiency, resulting in low generation efficiency and high labor cost of the project progress summary, but also is prone to errors (such as missing part of the progress information) of the professional personnel in sorting and summarizing each progress information due to the mixed progress information, resulting in low accuracy of the finally generated project progress summary. SUMMARY

[0003] The embodiments of the present application provide an information processing method, device, equipment and medium product, which can improve the generation efficiency and accuracy of the project progress summary and save labor cost.

[0004] In one aspect, the embodiments of the present application provide an information processing method, which comprises:

[0005] obtaining a plurality of progress record information of a sample media project, the sample media project being executed according to a media production process, one progress record information recording the execution progress of the sample media project at one time information, and the media production process defining a logical relationship between different execution progress;

[0006] calling a language model to summarize based on the plurality of progress record information of the sample media project, to obtain an initial project progress summary, the initial project progress summary including a plurality of clauses arranged in sequence, one clause being used to describe the execution progress of the sample media project at one time information;

[0007] based on the time information and the execution progress described by each clause in the initial project progress summary, calibrating the plurality of clauses in the initial project progress summary to obtain a target project progress summary; the plurality of clauses in the target project progress summary are arranged in chronological order, and the logical relationship between the execution progress described by the plurality of clauses matches the logical relationship defined by the media production process;

[0008] The language model is trained using the target project progress summary. The trained language model is then used to generate a corresponding media project progress summary based on multiple progress record information of any media project.

[0009] On the other hand, embodiments of this application provide an information processing apparatus, the apparatus comprising:

[0010] The acquisition unit is used to acquire multiple progress record information of the sample media project. The sample media project is executed according to the media production process. One progress record information records the execution progress of the sample media project under a certain time information. The media production process defines the logical relationship between different execution progress.

[0011] The processing unit is used to call the language model to summarize based on multiple progress record information of the sample media project to obtain an initial project progress summary. The initial project progress summary includes multiple clauses arranged in sequence, and one clause is used to describe the execution progress of the sample media project under a certain time information.

[0012] The processing unit is further configured to calibrate multiple clauses in the initial project progress summary based on the time information and execution progress described in each clause of the initial project progress summary to obtain a target project progress summary; the multiple clauses in the target project progress summary are arranged in chronological order, and the logical relationship between the execution progress described in the multiple clauses matches the logical relationship specified in the media production process;

[0013] The processing unit is further configured to train the language model using the target project progress summary, and the trained language model is used to generate a corresponding media project progress summary based on multiple progress record information of any media project.

[0014] In another aspect, embodiments of this application provide a computer device, the computer device including an input interface and an output interface, the computer device further including:

[0015] Processor and computer storage media;

[0016] The processor is adapted to implement one or more instructions, the computer storage medium stores one or more instructions, and the one or more instructions are adapted to be loaded by the processor and executed by the aforementioned information processing method.

[0017] In another aspect, embodiments of this application provide a computer storage medium storing one or more instructions, which are adapted to be loaded by a processor and executed by the aforementioned information processing method.

[0018] In another aspect, embodiments of this application provide a computer program product comprising one or more instructions; when one or more instructions in the computer program product are executed by a processor, they implement the information processing method mentioned above.

[0019] This application embodiment can acquire multiple progress record information of sample media projects executed according to the media production process, and call a language model to generate an initial project progress summary based on these multiple progress record information. Based on the time information and execution progress described in each clause of the initial project progress summary, the multiple clauses in the initial project progress summary are calibrated to obtain a target project progress summary. This ensures that the multiple clauses in the target project progress summary are arranged in chronological order, and that the logical relationships between the execution progress described in the multiple clauses match the logical relationships specified in the media production process. This guarantees that the target project progress summary has correct temporal and logical relationships. Therefore, when using the target project progress summary to train the language model, the language model can better understand the logical relationships between different execution progresses and improve its time series summarization ability during training. This enhances the logic and temporal rationality of the language model when generating project progress summaries. As a result, when the trained language model generates project progress summaries based on multiple progress record information of any media project, the risk of temporal confusion and logical errors in the project progress summary can be reduced, thereby improving the accuracy of the project progress summary. Furthermore, since the embodiments of this application automatically generate project progress summaries for any media project by training a language model and based on the trained language model, the embodiments of this application can realize the entire process of generating project progress summaries without the participation of professional personnel. This not only saves labor costs, reduces the risk of human error, and improves the accuracy of project progress summaries, but also avoids the problem of low efficiency in generating project progress summaries due to low human efficiency, thereby improving the efficiency of generating project progress summaries. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a logical schematic diagram of an information processing method provided in an embodiment of this application;

[0022] Figure 2 This is a flowchart illustrating an information processing method provided in an embodiment of this application;

[0023] Figure 3a This is a schematic diagram illustrating the generation of an initial project progress summary provided in an embodiment of this application;

[0024] Figure 3b This is a schematic diagram illustrating another method for generating an initial project progress summary, provided in an embodiment of this application.

[0025] Figure 3c This is a schematic diagram of a process for calibrating an initial project progress summary to obtain a target project progress summary, as provided in an embodiment of this application.

[0026] Figure 4 This is a flowchart illustrating an information processing method provided in another embodiment of this application;

[0027] Figure 5a This is a schematic diagram of a progress recording information provided in an embodiment of this application;

[0028] Figure 5b This is a schematic diagram illustrating an early warning system for an initial project progress summary, provided in an embodiment of this application.

[0029] Figure 5c This is a schematic diagram of the calculation logic for expressing loss values ​​provided in an embodiment of this application;

[0030] Figure 5d This is a schematic diagram illustrating the calculation logic of a ranking loss value provided in an embodiment of this application;

[0031] Figure 5e This is a schematic diagram of the training logic of a language model provided in an embodiment of this application;

[0032] Figure 6a This is a schematic diagram illustrating the application process of a language model provided in an embodiment of this application;

[0033] Figure 6b This is a schematic diagram illustrating an archiving result provided in an embodiment of this application;

[0034] Figure 6c This is a schematic diagram illustrating a project progress summary provided in an embodiment of this application;

[0035] Figure 7 This is a schematic diagram of the structure of an information processing device provided in an embodiment of this application;

[0036] Figure 8 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0037] To facilitate understanding of the specific implementation of the technical solutions in the embodiments of this application, the key technical terms involved in the embodiments of this application are introduced below:

[0038] I. Media Projects

[0039] Media projects, also known as media production projects, can include, but are not limited to, any of the following: film and television production projects, audio production projects, animation production projects, advertising production projects, etc. Film and television production projects refer to projects that produce film and television content, such as movies, TV series, web series, and short videos (videos with a playback length less than a certain threshold). Audio production projects refer to projects that produce audio content, such as musical works and radio dramas. Animation production projects refer to projects that produce animation, such as 2D animation, 3D animation, and stop-motion animation. Advertising production projects refer to projects that produce advertising content, such as television commercials (advertisements broadcast on television) and online advertising (advertisements broadcast on internet platforms).

[0040] II. Media Production Process

[0041] The media production process refers to a series of processes from the initiation of a media project to its completion. It may include multiple sequentially progressing project phases, each involving one or more steps, with overlapping steps between phases. For example, taking a film or television production project as an example, the media production process may include the following phases: intent confirmation phase, scriptwriting phase, contract preparation phase, filming phase, post-production phase, review and approval phase, and distribution and promotion phase. Specifically, the intent confirmation phase may involve information gathering and determining the film / television production intent; the scriptwriting phase may involve searching for or creating a script (i.e., creating the initial draft), revising the script, and finalizing the script; the contract preparation phase may involve signing contracts with the production company and preparing materials; the filming phase may involve filming and revising the script; the post-production phase may involve editing various versions, adding music, and correcting edited content; the review and approval phase may involve multiple parties reviewing and revising content, submitting multiple versions for review, and correcting edited content; and the distribution and promotion phase may involve publicity and promotion, correcting edited content, agreeing on broadcast formats, negotiating scheduling, setting a release date, launching the program, and completing the broadcast. It is evident that in this media production process, steps such as revising the script and correcting the editing can be carried out throughout at least two project stages.

[0042] III. Project Progress Summary

[0043] A project progress summary is a type of information used to summarize the progress of a media project. It not only indicates the current stage of the project and the execution status of that stage (e.g., execution progress, estimated completion time), but also the execution status of completed stages (e.g., whether each stage was completed on time, whether there were any delays or early completions, and any problems encountered during the execution of each stage). Specifically, a project progress summary refers to the information obtained by summarizing (e.g., analyzing, inductively combining, and refining) multiple progress records of a media project. These progress records are generated by various stakeholders in the media project, recording the execution progress of the project at a specific point in time. These stakeholders refer to the personnel, organizations, or platforms involved in the execution of the media project. Taking a film or television production project as an example, these stakeholders may include, but are not limited to, the original author, director, screenwriter, contracting platform, review body, broadcasting organization, and platform operation and follow-up personnel.

[0044] Based on the above introduction of key technical terms, the technical solutions of the embodiments of this application will be clearly and completely described below.

[0045] This application proposes an information processing method that can automatically generate project progress summaries for media projects by training a language model. The entire generation process of the project progress summary does not require professional personnel, thus saving labor costs, reducing the risk of human error, improving the accuracy of the project progress summary, and avoiding the problem of low efficiency in project progress summary generation due to low manual efficiency. The language model, also known as a large language model (LLM), is an artificial intelligence model built based on deep learning technology, composed of artificial neural networks with hundreds of millions of parameters.

[0046] See Figure 1As shown, the general principle of the information processing method proposed in this application embodiment is as follows: Multiple progress record information of a sample media project executed according to the media production process is obtained. This progress record information can originate from various personnel within the sample media project. A language model is invoked to generate an initial project progress summary based on this multiple progress record information. This initial project progress summary includes multiple clauses arranged sequentially, with each clause describing the execution progress of the sample media project at a given time. Further, based on the time information and execution progress described by each clause in the initial project progress summary, the multiple clauses in the initial project progress summary are calibrated to obtain a target project progress summary. The multiple clauses in the target project progress summary are arranged in chronological order, and the logical relationship between the execution progress described by the multiple clauses matches the logical relationship specified in the media production process. Furthermore, the target project progress summary can be used to train the language model, enabling the trained language model to understand the logical relationship between different execution progresses and to summarize time series data. This allows the trained language model to generate a project progress summary with correct temporal and logical relationships based on multiple progress record information of any media project.

[0047] In specific implementations, the information processing method proposed in this application can be executed by a computer device, which can be a terminal or a server; that is, the method can be executed by the terminal or the server alone, or by both the terminal and the server together, without limitation. The terminal can be a smartphone, computer (such as a tablet, laptop, or desktop computer), smart wearable device (such as a smartwatch or smart glasses), smart voice interaction device, smart home appliance (such as a smart TV), vehicle terminal, or aircraft, etc.; the server can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms, etc.

[0048] It is worth emphasizing that in the embodiments of this application, if user information and other related data are involved, when any method embodiment proposed in the embodiments of this application is applied to a specific product or technology, such related data is collected with the user's permission or consent, and the collection, use and processing of such data comply with the relevant laws, regulations and standards of the relevant regions.

[0049] Based on the above description, the following section uses a computer device as the execution subject as an example, combined with... Figure 2The flowchart shown illustrates the specific implementation process of the information processing method proposed in the embodiments of this application. See also... Figure 2 As shown, the method can generally include the following steps S201-S204:

[0050] S201, Obtain multiple progress record information for the sample media project.

[0051] Sample media projects refer to media projects used for training the language model. Specifically, they can be any film and television production project, any audio production project, any animation production project, any advertising production project, etc., without any limitation.

[0052] The sample media project is executed according to a media production process, which may include multiple project phases that proceed sequentially. A progress record for the sample media project documents its execution progress at a specific point in time, specifically the progress of the project phase at that point. The execution progress of a project phase refers to the status of the sample media project within that phase. For example, if the project phase is scriptwriting, its progress could be "script revision" or "script completed." Similarly, if the project phase is filming, its progress could be "filming in progress," "filming 30% complete," or "filming completed." And if the project phase is submission for review, its progress could be "submission completed."

[0053] Understandably, each phase of a sample media project can generate one or more execution schedules. The media production workflow defines the sequence of project phases, indicating the logical relationships such as the order of execution and dependencies between them. Therefore, the execution schedules of different project phases should also follow the logical relationships between those phases. Thus, the media production workflow defines the logical relationships between project phases through the sequence of their execution, thereby defining the logical relationships between different execution schedules. These logical relationships may include, but are not limited to, at least one of the following: the order of execution schedules, dependencies between execution schedules, etc.

[0054] The sequence of execution progress is used to indicate the order in which execution progress is arranged, such as the order of "completing the script first, then filming". The dependency relationship between execution progress is used to indicate whether there is a dependency between any two execution progress. When the generation of one execution progress must be a prerequisite for the generation of another execution progress, it can be considered that the execution progress depends on the other execution progress. For example, the execution progress "in post-production editing" can only be generated after the execution progress "filming completed" is generated. Therefore, the execution progress "in post-production editing" can be considered to depend on the execution progress "filming completed".

[0055] S202, the language model is invoked to summarize the progress information of multiple progress records of the sample media project, and an initial project progress summary is obtained.

[0056] In practical implementation, the input text for the language model can be constructed using multiple progress record information from the sample media project. This input text includes multiple progress record information from the sample media project. Furthermore, this input text can be input into the language model, which then generates an initial project progress summary based on this input text. Taking an example where the sample media project has H progress record information (H is an integer greater than 1), a schematic diagram illustrating the generation of the initial project progress summary in this case can be found here. Figure 3a As shown. Since the initial project progress summary is generated by the language model, it can also be called the model project progress summary or the machine project progress summary. The initial project progress summary includes multiple clauses arranged in sequence, with each clause describing the execution progress of the sample media project under a certain time information.

[0057] The language model can be configured with a dictionary containing multiple characters, where a character can be a Chinese character, letter, number, operator, punctuation mark, or other symbol. Based on this, a specific implementation method for the language model to generate an initial project progress summary from the input text can be exemplarily described in [reference needed]. Figure 3b As shown, it is roughly as follows: Based on the input text, the probability of each character in the dictionary is predicted multiple times. Each time a prediction is made, the character with the highest probability is selected from the dictionary as the currently generated character based on the probability of each predicted character. The currently generated character is added to the input information to update the input information, thereby triggering the next prediction, until the multiple predictions are completed, and the initial project progress summary is obtained (composed of each character generated during the multiple predictions).

[0058] S203, based on the time information and execution progress described in each clause of the initial project progress summary, calibrate multiple clauses in the initial project progress summary to obtain the target project progress summary.

[0059] In one specific implementation, the computer device can directly rearrange the clauses in the initial project progress summary according to the time information and execution progress described by each clause, so as to calibrate the arrangement order between the clauses in the initial project progress summary and obtain the target project progress summary. The multiple clauses in the target project progress summary are arranged in chronological order, and the logical relationship between the execution progress described by the multiple clauses matches the logical relationship specified in the media production process.

[0060] The meaning of "matching the logical relationship between the execution progress described by multiple clauses with the logical relationship stipulated in the media production process" can be: the logical relationship between the execution progress described by multiple clauses is the same as (consistent with) the logical relationship stipulated in the media production process, or the logical relationship between the execution progress described by multiple clauses is part of the logical relationship stipulated in the media production process. For example, the logical relationship stipulated in the media production process includes: finalized script → filming → post-production → submission for review; if multiple clauses include clause 1, clause 2, and clause 3, and clause 1 describes the execution progress as "the script has been finalized", clause 2 describes the execution progress as "filming progress 30%", and clause 3 describes the execution progress as "post-production", then the logical relationship between the execution progress described by these three clauses is: finalized script → filming → post-production; in this case, it can be determined that the logical relationship between the execution progress described by these three clauses matches the logical relationship stipulated in the media production process.

[0061] In another specific implementation, the computer equipment can also calibrate the descriptions of each clause in the initial project progress summary based on summary expression guidelines in the media production field to obtain an intermediate project progress summary, ensuring that the descriptions of each clause in the intermediate project progress summary conform to the summary expression guidelines. Furthermore, in the intermediate project progress summary, the clauses can be rearranged according to the time information and execution progress described in each clause to calibrate the order of the clauses in the intermediate project progress summary, thus obtaining the target project progress summary. In this case, the general process of calibrating the initial project progress summary to obtain the target project progress summary can be exemplarily described in [reference needed]. Figure 3c As shown, in this case, the multiple clauses in the target project progress summary are arranged in chronological order, and the logical relationship between the execution progress described by the multiple clauses matches the logical relationship specified in the media production process. Furthermore, the expression of each clause conforms to the summary expression guidelines in the field of media production.

[0062] The summary and expression guidelines in the media production field are constructed based on domain knowledge within that field. When the media production field is film and television production, these summary and expression rules can also be referred to as film and television industry summary management rules. For example, summary and expression rules may include, but are not limited to, at least one of the following: professional terminology involved in each project stage of the media production process, and language expression standards for the execution progress of each project stage (such as language expression format, language style, etc.). It is evident that by introducing summary and expression rules to calibrate the initial project progress summary, the expression of each clause in the calibrated target project progress summary conforms to the summary and expression guidelines in the media production field, avoiding vague or inappropriate descriptions in the clauses of the target project progress summary. By using such a target project progress summary to train the language model, the language model can accurately learn the summary and expression guidelines and domain knowledge in the media production field based on the target project progress summary, achieving the embedding of domain knowledge into the language model. This reduces vague or inappropriate descriptions output by the language model in project progress summary tasks, thereby improving the accuracy of the final project progress summary generated by the language model.

[0063] S204, The language model is trained using the target project progress summary. The trained language model is used to generate the corresponding media project progress summary based on multiple progress record information of any media project.

[0064] In one specific implementation, the computer device can calculate the model loss value of the language model based on the target project progress summary, and then optimize the model parameters of the language model based on the model loss value. Specifically, the model loss value may include at least one of the following: correctness loss value, expression loss value, and ranking loss value. Among them: (1) the correctness loss value refers to the loss value generated by the language model in the dimension of summary correctness, which can be used to measure the correctness of the project progress summary generated by the language model. The larger the correctness loss value, the lower the correctness of the project progress summary generated by the language model; (2) the expression loss value refers to the loss value generated by the language model in the dimension of summary expression rationality, which can be used to measure the expression rationality of each clause in the project progress summary generated by the language model. The larger the expression loss value, the lower the expression rationality of each clause in the project progress summary; (3) the ranking loss value refers to the loss value generated by the language model in the dimension of clause ranking, which can be used to measure the ranking correctness of each clause in the project progress summary generated by the language model. The larger the ranking loss value, the lower the ranking correctness of each clause in the project progress summary.

[0065] In another specific implementation, the computer device can also use the target project progress summary as a label for the sample media project, and then optimize the model parameters of the language model based on this label and the initial project progress summary. For example, the similarity between the label and the initial project progress summary can be calculated, and the model parameters of the language model can be optimized in the direction of increasing similarity. Here, "in the direction of increasing similarity" means optimizing the model in a way that maximizes similarity. Optimizing the model in this direction ensures that the similarity between the initial project progress summary and the label generated by the language model after each optimization is greater than the similarity between the initial project progress summary and the label generated by the language model before optimization. For example, if the similarity between the initial project progress summary and the label calculated in this case is 0.5, then after optimizing the model parameters of the language model in the direction of increasing similarity, the similarity between the initial project progress summary and the label generated by the optimized language model should be greater than 0.5.

[0066] This application embodiment can acquire multiple progress record information of sample media projects executed according to the media production process, and call a language model to generate an initial project progress summary based on these multiple progress record information. Based on the time information and execution progress described in each clause of the initial project progress summary, the multiple clauses in the initial project progress summary are calibrated to obtain a target project progress summary. This ensures that the multiple clauses in the target project progress summary are arranged in chronological order, and that the logical relationships between the execution progress described in the multiple clauses match the logical relationships specified in the media production process. This guarantees that the target project progress summary has correct temporal and logical relationships. Therefore, when using the target project progress summary to train the language model, the language model can better understand the logical relationships between different execution progresses and improve its time series summarization ability during training. This enhances the logic and temporal rationality of the language model when generating project progress summaries. As a result, when the trained language model generates project progress summaries based on multiple progress record information of any media project, the risk of temporal confusion and logical errors in the project progress summary can be reduced, thereby improving the accuracy of the project progress summary. Furthermore, since the embodiments of this application automatically generate project progress summaries for any media project by training a language model and based on the trained language model, the embodiments of this application can realize the entire process of generating project progress summaries without the participation of professional personnel. This not only saves labor costs, reduces the risk of human error, and improves the accuracy of project progress summaries, but also avoids the problem of low efficiency in generating project progress summaries due to low human efficiency, thereby improving the efficiency of generating project progress summaries.

[0067] Based on the above description, this application proposes another information processing method; in this application embodiment, the method is still described using a computer device executing the information processing method as an example. Please refer to... Figure 4 As shown, this information processing method can generally include the following steps S401-S406:

[0068] S401, retrieve multiple progress record information for the sample media project.

[0069] In its implementation, the sample media project can be executed according to a media production process, which may include multiple project stages. During the execution of the sample media project, all parties involved (such as the original author, director, screenwriter, contracting platform, review agency, and operations personnel) can record the project's progress at at least one time point and upload this progress record information to a database via a web application, mobile application, or a designated interface. When the computer device executes step S401, it can retrieve multiple progress record information entries for the sample media project from this database. Each progress record entry represents the project's execution progress at a given time point.

[0070] The information format of each progress record can be as follows: Figure 5a As shown. Alternatively, each progress record can also use a fixed description A (such as a standard time template: year x month x day, execution progress x; or a progress node template: year x month x day, x progress is x%) + a variable description B (custom remarks, special event descriptions) to ensure that key information (such as time information, information used to indicate execution progress) is structured and that special information is flexibly supplemented. For example, to facilitate subsequent traceability, each progress record can be automatically marked with custom remarks such as the entry time (time uploaded to the database), the responsible person, the project identifier, and stage tags. Among them, stage tags refer to the tags of project stages, such as the tags for scriptwriting, production company confirmation, filming, post-production, submission for review, broadcast scheduling, premiere, and completion of broadcast. The specific stage tags can be selected by the person inputting the progress record information.

[0071] As can be seen, in this embodiment, each progress record can include a fixed description field and a variable description field. The fixed description field may include time information and stage progress information (such as the aforementioned execution progress x or x% progress), which indicates the execution progress of the sample media project at the corresponding time stage. The variable description field may include remarks related to the stage progress information, which may include, but are not limited to, descriptions of special events and custom remarks. For example, a specific example of progress record information can be as follows:

[0072] Submitted for review on May 30, 2025, awaiting confirmation. Delayed due to holiday; confirmation will be provided after the holiday. (Note: Updated June 1, 2025, Operations Manager Zhang Yier, Project 2, submitted for review).

[0073] As of May 29, 2025, filming progress was 24%. Set construction at Location A expires on June 20, and 30% of filming must be completed before the expiration date. (Note: Updated June 1, 2025, Director: Zhang Sansan, Project 100, Filming).

[0074] S402, call the language model to summarize based on multiple progress record information of the sample media project to obtain the initial project progress summary.

[0075] This application does not limit the specific selection of the language model or the model structure. For example, the language model can use a dictionary of a preset length (e.g., 152064) and generate an embedding of a preset size (e.g., 1×5120) for each character in the dictionary to form the vector space of the dictionary. The language model can use a 64-layer decoder layer as the main structure of the model to decode the input data. A normalization layer is connected after the main structure of the model. The normalization layer generates the embedding based on the decoding result output by the main structure of the model. An lm_head structure (linear layer or fully connected layer) is connected after the normalization layer. The lm_head structure maps the embedding generated by the normalization layer to the vector space corresponding to the dictionary to predict the probability of each character in the dictionary, and then selects the character for output based on the probability prediction result.

[0076] Based on this, the specific implementation of step S402 can be as follows: The input text of the language model is generated using multiple progress record information from the sample media project; the input information is segmented into multiple characters, and each character is encoded into a vector to obtain the text vector corresponding to the input text; the 64-layer decoder layer in the language model is called to decode the text vector, and the normalization layer is called to process the decoding result to obtain the embedding of the input text; the lm-head structure is called to predict the probability of each character in the dictionary based on the embedding of the input text. Specifically, the embedding of the input text can be mapped to the vector space of the dictionary, and the vector similarity between the embedding of the input text and the embedding of each character can be calculated in this vector space. The probability of each character in the dictionary can be predicted based on the calculated vector similarity, thereby selecting the character with the highest probability as the first predicted character. Furthermore, the vector of the first predicted character can be fused into the text vector of the input text to obtain a new text vector. The language model is then called to predict the probability of each character in the dictionary again based on the new text vector according to the above logic. The character with the highest probability is selected as the second predicted character, and so on. After obtaining Q predicted characters, an initial project progress summary can be obtained. This initial project progress summary includes Q predicted characters, where Q is an integer greater than 1, and the Q predicted characters are divided into multiple clauses arranged in sequence. Each clause is used to describe the execution progress of the sample media project under a certain time information.

[0077] It should be noted that the above description is the specific implementation process of the language model summarizing multiple progress records of the sample media project to obtain an initial project progress summary. In the specific implementation of step S402, considering that problems such as ambiguity and incorrect order of clauses may occur when the language model generates a project progress summary, which often recur in multiple predictions, the language model can perform N summaries based on multiple progress records of the sample media project to obtain N initial project progress summaries (each summary yields one initial project progress summary, where N is a positive integer). These N initial project progress summaries are used to capture these problems, so that they can be addressed specifically during subsequent language model training, thereby improving the accuracy of the language model in generating project progress summaries.

[0078] Optionally, after obtaining the initial project progress summary, the computer device can also call the language model to perform uncertainty detection on each clause in the initial project progress summary and obtain the detection results. The uncertainty of a clause is used to indicate the credibility of the language model for that clause; when a clause has uncertainty, it indicates that the language model has low credibility for that clause. For example, a specific implementation method for the language model to perform uncertainty detection on each clause in the initial project progress summary can be: performing conflict detection on the time information described by any two adjacent clauses in the initial project progress summary, that is, detecting whether the order of the time information described by any two adjacent clauses satisfies chronological order, and detecting whether the logical relationship between the execution progress described by any two adjacent clauses matches the logical relationship specified in the media production process; if the order of the time information described by two adjacent clauses does not satisfy chronological order, it can be determined that these two adjacent clauses have failed the conflict detection, and it can be determined that these two adjacent clauses have uncertainty; if the logical relationship between the execution progress described by two adjacent clauses does not match the logical relationship specified in the media production process, it can be determined that these two adjacent clauses have uncertainty.

[0079] Uncertainty detection is performed on each clause in the initial project progress summary. If the detection result indicates that there is an uncertain clause in the initial project progress summary, an early warning is issued. This application embodiment does not limit the warning method; for example, a prompt message can be output to indicate that at least one clause in the initial project progress summary has uncertainty, or the initial project progress summary can be output on the terminal interface, and clauses with uncertainty can be visually highlighted (e.g., highlighted, magnified, etc.). See, for example... Figure 5b As shown, the second clause in the initial project progress summary is "The project was started on [Date], and the platform is paying close attention to the progress," while the fourth clause is "The project was started on [Date], and the filming progress has reached 60%." Since the time information described in these two clauses conflicts, they can be considered to contain uncertainty. Therefore, when displaying the initial project progress summary, these two clauses can be highlighted. It is evident that this embodiment can more easily detect and issue warnings when the language model detects clauses with uncertainty (such as "contradictory start-up times"), facilitating manual intervention. This risk perception and warning system allows relevant personnel to easily track project progress and improves project management efficiency.

[0080] S403, based on the summary and expression guidelines in the field of media production, calibrates the expression of each clause in the initial project progress summary to obtain the intermediate project progress summary.

[0081] As mentioned above, there are N initial project progress summaries. In the specific implementation of step S403, the computer equipment can calibrate the expression of each clause in each initial project progress summary based on the summary expression criteria in the field of media production, and obtain N intermediate project progress summaries.

[0082] One specific implementation method for using computer equipment to calibrate the wording of each clause in an initial project progress summary based on summary expression guidelines in the media production field, to obtain an intermediate project progress summary, could be: 1) Send the initial project progress summary to professionals, who then manually calibrate the wording of each clause in the initial project progress summary based on summary expression guidelines in the media production field, thus obtaining the intermediate project progress summary. 2) Call a pre-trained expression calibration model to calibrate the wording of each clause in the initial project progress summary based on summary expression guidelines in the media production field, thus obtaining the intermediate project progress summary. 3) Based on summary expression guidelines in the media production field, determine a manual project progress summary for a sample media project (i.e., a project progress summary generated by professionals based on multiple progress records of the sample media project and conforming to summary expression guidelines in the media production field), and then calibrate the wording of each clause in the initial project progress summary based on the wording of each clause in this manual project progress summary, thus obtaining the intermediate project progress summary, and so on.

[0083] S404, in the intermediate project progress summary, the clauses are rearranged according to the time information and execution progress described in each clause to obtain the target project progress summary.

[0084] As mentioned above, there are N intermediate project progress summaries. In the specific implementation of step S404, the computer device can rearrange each clause in each intermediate project progress summary according to the time information and execution progress described by each clause to obtain N target project progress summaries. The multiple clauses in each target project progress summary are arranged in chronological order, and the logical relationship between the execution progress described by the multiple clauses matches the logical relationship specified in the media production process.

[0085] One specific implementation method for a computer device to rearrange the clauses in an intermediate project progress summary based on the time information and execution progress described in each clause, thus obtaining a target project progress summary, could be as follows: The intermediate project progress summary is sent to professionals, who, based on the media production workflow, manually rearrange the clauses in the received intermediate project progress summary according to the time information and execution progress described in each clause, thus obtaining the target project progress summary. Alternatively, a pre-trained ranking model can be used to rearrange the clauses in the intermediate project progress summary based on the media production workflow and the time information and execution progress described in each clause, thus obtaining the target project progress summary, and so on.

[0086] S405, based on the summary of the target project progress, calculate the model loss value of the language model.

[0087] The model loss value includes at least one of the following: correctness loss, representation loss, and ranking loss. Specifically:

[0088] (1) When the model loss value includes the correctness loss value, as mentioned above, the number of target project progress summaries is N. In this case, the specific implementation of step S405 may include the following steps s11-s14:

[0089] s11. Select one target project progress summary as the correct project progress summary from N target project progress summaries. Specifically, one target project progress summary can be randomly selected from the N target project progress summaries, or the optimal target project progress summary can be selected from the N target project progress summaries. The optimal target project progress summary can be determined manually by sending the N target project progress summaries to professionals for comparison to determine the optimal one; alternatively, the optimal target project progress summary can be determined using a model comparison method, where a pre-trained quality detection model is used to perform quality checks on each of the N target project progress summaries, and the optimal target project progress summary is determined based on the quality detection results; or, the optimal target project progress summary can also be determined by comparing the manual project progress summaries, where the difference between each target project progress summary and the manual project progress summary is calculated, and the target project progress summary with the smallest difference is determined as the optimal target project progress summary.

[0090] s12, assign weights to each clause in the correct project schedule summary to obtain a weight score for each clause. This weight score is negatively correlated with the clause's position in the correct project schedule summary. Specifically, based on the position of each clause in the correct project schedule summary, a relevance score (rel) can be defined for each clause. This relevance score is related to the clause's position and can be negatively correlated with it; that is, the later the clause is in the sequence, the smaller its relevance score. For example, if there are I clauses in the correct project schedule summary (I is an integer greater than 1), the relevance scores of each clause can be I, I-1, ..., 1. Furthermore, the relevance score of each clause can be directly used as the weight score of the corresponding clause, or a score reduction factor can be obtained for each clause. This factor can be used to reduce the relevance score of the corresponding clause to obtain its weight score. For example, the relevance score of the i-th clause is rel. i And if the score reduction factor of the i-th clause can be log2(i+1), then the weight score of the i-th clause can be rel. i / log2(i+1).

[0091] s13, Obtain the generation probability of each character in the correct project progress summary. This generation probability refers to the probability that the language model generates the corresponding character. Specifically, the language model can be guided to generate the correct project progress summary based on multiple progress record information of the sample media project to obtain the generation probability of each character in the correct project progress summary; or, if the correct project progress summary is obtained by calibrating the initial project progress summary generated by the language model, then the generation probability of each character in the correct project progress summary can be determined based on the generation probability of each character in the initial project progress summary. For example, if the first character in the correct project progress summary is obtained by calibrating the first character in the initial project progress summary, then the generation probability of the first character in the initial project progress summary can be used as the generation probability of the first character in the correct project progress summary. Understandably, as mentioned above, when generating any project progress summary, the language model can predict the probability of each character in the dictionary, thereby generating the character with the highest probability to construct the project progress summary. Therefore, the generation probability of any character can be a probability distribution, which can include the predicted probability of each character in the dictionary. For example, if the dictionary contains 152,064 characters, then the generation probability of any character can be a 1×152,064-dimensional probability distribution, which includes the predicted probabilities of 152,064 characters.

[0092] s14. Based on the generation probability of each character in the correct project progress summary and the weight score of each clause, calculate the correctness loss value of the language model. Specifically, it is possible to traverse each clause in the correct project progress summary, and calculate the loss value of the i-th clause based on the generation probability of each character in the currently traversed i-th clause and the weight score of the i-th clause, where i∈[1,P], and P is the number of clauses in the correct project progress summary; continue to traverse each clause until the loss value of each clause in the correct project progress summary is obtained, and integrate the loss values ​​of each clause (such as by mean calculation) to obtain the correctness loss value of the language model. The specific implementation of calculating the loss value of the i-th clause based on the generation probability of each character in the i-th clause and the weight score of the i-th clause can be as follows: perform a logarithmic operation (log) on ​​the generation probability of each character in the i-th clause to obtain the logarithmic operation result of each character in the i-th clause; in addition, the label vector of each character in the i-th clause can be obtained. The label vector of each character is determined based on the number of characters contained in the dictionary and the arrangement position of the corresponding character in the dictionary. For example, if the dictionary contains 152064 characters, and the c-th character in the i-th clause (c∈[1,M], where M is the number of characters in the i-th clause) is arranged at position 50 in the dictionary, then the label vector of the c-th character is a 1×152064-dimensional vector, and the 50th element in the vector is 1, and all other elements are 0. Furthermore, the loss value of the c-th character in the i-th clause can be calculated using the logarithmic result of the c-th character in the i-th clause, the label vector, and the weight score of the i-th clause. After obtaining the loss values ​​of each character in the i-th clause, the loss values ​​of each character in the i-th clause can be summed to obtain the loss value of the i-th clause.

[0093] Based on the description of steps s11-s14 above, one way to calculate the correctness loss value of the language model can be exemplarily shown in the following formula 1.1:

[0094] Loss1=-1 / P*sum_i(sum_c(yic*log(pic)*w_nDCG)) Formula 1.1

[0095] In Formula 1.1 above, Loss1 represents the correctness loss value, w_nDCG represents the weight score of the i-th clause, pic represents the generation probability of the c-th character in the i-th clause, yic represents the label vector of the c-th character in the i-th clause, and sum represents summation.

[0096] It should be noted that steps s11-s14 above are merely illustrative descriptions of one implementation method for calculating the correctness loss value of a computer device's language model, and are not intended to limit the scope of the implementation. For example, in other embodiments, the cross-entropy loss function can also be used to calculate the correctness loss value of the language model based on the generation probability of each character in the correct project progress summary. In this case, the calculation method for the correctness loss value can be exemplarily shown in Formula 1.2 below:

[0097]

[0098] In Formula 1.2 above, Li represents the loss value of the i-th clause in the correct project progress summary, while the definitions of other letters are the same as those in Formula 1.1, and will not be repeated here. By comparing Formula 1.1 and Formula 1.2, it can be seen that, compared to Formula 1.2, Formula 1.1 adds the weight score of the clause in which the corresponding character is located to each character in the correct project progress summary to control the loss value of the corresponding character. This can improve the accuracy of the correctness loss value of the final calculated language model.

[0099] (2) When the model loss value includes the expression loss value, as mentioned above, the target project progress summary is obtained by calibrating the order of the clauses in the intermediate project progress summary, and the intermediate project progress summary is obtained by calibrating the expression of the clauses in the initial project progress summary. In this case, the specific implementation of step S405 may include the following steps s21-s23:

[0100] s21, using the target project progress summary, intermediate project progress summary, and initial project progress summary, multiple original sample pairs are constructed. Each original sample pair includes positive and negative samples. The positive sample in any original sample pair is the target project progress summary, and the negative sample in any original sample pair is either the intermediate project progress summary or the initial project progress summary. The negative samples in different original sample pairs are different. As mentioned above, the language model can perform N summaries based on multiple progress record information of the sample media project to obtain N initial project progress summaries. Then, when calibrating multiple clauses in the nth initial project progress summary, the nth intermediate project progress summary and the nth target project progress summary can be obtained, n∈[1,N]. In this case, the nth target project progress summary and the nth initial project progress summary construct one original sample pair, and the nth target project progress summary and the nth intermediate project progress summary construct one original sample pair. Therefore, by calibrating an initial project progress summary, two original sample pairs can be constructed. Thus, by calibrating N initial project progress summaries, 2N original sample pairs can be obtained.

[0101] s22, obtain the positive example generation probability and negative example generation probability for each original sample pair. The positive example generation probability refers to the probability that the language model generates a positive sample, which can be calculated based on the generation probability of each character in the positive sample (i.e., the probability that the language model generates the corresponding character). Similarly, the negative example generation probability refers to the probability that the language model generates a negative sample, which can be calculated based on the generation probability of each character in the negative sample (i.e., the probability that the language model generates the corresponding character).

[0102] s23. Based on the positive and negative example generation probabilities of each original sample pair, calculate the expression loss value of the language model. Specifically, the difference between the positive and negative example generation probabilities of each original sample pair can be calculated to obtain the loss value of each original sample pair. These loss values ​​are then integrated to obtain the expression loss value of the language model. For example, using a target project progress summary, a schematic diagram illustrating the calculation logic of this expression loss value can be found in [reference needed]. Figure 5c As shown in the diagram. The difference between the probability of generating a positive example and the probability of generating a negative example can be the difference between the probability of generating a positive example and the probability of generating a negative example, or the difference between these two probabilities after performing a logarithmic operation. That is, the first result can be obtained by performing a logarithmic operation on the probability of generating a positive example, and the second result can be obtained by performing a logarithmic operation on the probability of generating a negative example. The difference between the first result and the second result is taken as the difference between the probability of generating a positive example and the probability of generating a negative example.

[0103] Alternatively, a specific implementation method for calculating the expression loss value of the language model based on the positive and negative example generation probabilities of each original sample pair can be: obtaining a first reference probability and a second reference probability for each original sample pair, where the first reference probability is the probability that the reference model generates a positive sample in the original sample pair, and the second reference probability is the probability that the reference model generates a negative sample in the original sample pair; wherein, the language model can be trained once or multiple times, and the reference model here refers to the model obtained by copying the language model before training the language model, which is not updated during the training process of the language model. Furthermore, for any given original sample pair, the positive example loss value of the original sample pair can be determined based on the ratio between the positive example generation probability and the first reference probability, and the negative example loss value of the original sample pair can be determined based on the ratio between the negative example generation probability and the second reference probability. Thus, the difference between the positive and negative example loss values ​​of the original sample pair can be calculated to obtain the loss value of the original sample pair. In addition, the loss values ​​of each original sample pair can be integrated to obtain the expression loss value of the language model.

[0104] One specific implementation method for determining the positive example loss value of the original sample pair based on the ratio between the positive example generation probability and the first reference probability of the original sample pair can be: directly using the ratio between the positive example generation probability and the first reference probability of the original sample pair as the positive example loss value of the original sample pair; or performing a logarithmic operation on the ratio between the positive example generation probability and the first reference probability of the original sample pair to obtain the positive example loss value of the original sample pair; or performing a logarithmic operation on the ratio between the positive example generation probability and the first reference probability of the original sample pair, obtaining a first operation result, and then scaling the first operation result using a scaling factor to obtain the positive example loss value of the original sample pair, and so on. Similarly, the specific methods for determining the negative example loss value of the original sample pair based on the ratio between the negative example generation probability and the second reference probability can include: directly using the ratio between the negative example generation probability and the second reference probability of the original sample pair as the negative example loss value of the original sample pair; performing a logarithmic operation on the ratio between the negative example generation probability and the second reference probability of the original sample pair to obtain the negative example loss value of the original sample pair; performing a logarithmic operation on the ratio between the negative example generation probability and the second reference probability of the original sample pair to obtain the second operation result, and then scaling the second operation result using a scaling factor to obtain the negative example loss value of the original sample pair, and so on.

[0105] Based on the description of steps s21-s23 above, one way to calculate the expressive loss value of the language model can be exemplarily shown in the following formula 2.1:

[0106]

[0107] In Formula 2.1 above, Loss2 represents the loss value. This represents the expectation operator, where x represents the input text constructed based on multiple progress records from the sample media project, and y represents the expectation operator. w Let y represent a positive sample. L Let represent negative samples, D represent the set containing all original sample pairs, σ be a sigmoid function (activation function), and π be a negative sample. θ (y w |x) represents the positive example generation probability of the original sample pair, π ref (y w |x) represents the first reference probability of the original sample pair, π θ (y L |x) represents the probability of generating a negative example of the original sample pair, π ref (y L |x) represents the second reference probability of the original sample pair. This represents the positive example loss value of the original sample pair. β represents the negative loss value of the original sample pair, and β represents the scaling factor, which can be a hyperparameter and its value can be set based on empirical values.

[0108] It should be noted that steps s21-s23 described above are merely illustrative of one implementation method for a computer device to calculate the expressive loss value of a language model, and are not intended to limit the scope of the implementation. For example, in other embodiments, the computer device may also construct multiple original sample pairs based on step s21, calculate the sample difference between positive and negative samples in each original sample pair, obtain the loss value of each original sample pair, and then integrate the loss values ​​of each original sample pair to obtain the expressive loss value of the language model. Furthermore, in other embodiments, the target project progress summary may be obtained by directly calibrating the order of the clauses in the initial project progress summary. In this case, there is no intermediate project progress summary, so the computer device can construct multiple original sample pairs using only N initial project progress summaries and corresponding N target project progress summaries, and then perform steps s22-s23 based on these multiple original sample pairs to obtain the expressive loss value of the language model.

[0109] (3) When the model loss value includes the ranking loss value, as mentioned above, the number of target project progress summaries is N. A target project progress summary can be obtained by directly calibrating an initial project progress summary, or it can be obtained by calibrating an intermediate project progress summary after calibrating the representation of an initial project progress summary. There is no limitation on this. In this case, the specific implementation of step S405 may include the following steps s31-s34:

[0110] s31. Select one target project progress summary as the correct project progress summary from N target project progress summaries. The specific implementation method of this step can be found in the specific implementation method of step s11 above, and will not be repeated here.

[0111] S32 involves rearranging the clauses in the correct project progress summary K times to obtain K incorrect project progress summaries. Each rearrangement generates one incorrect project progress summary. Any two incorrect project progress summaries have different clause orders, where K is a positive integer. For example, if the correct project progress summary includes 3 clauses, namely clause A, clause B, and clause C, then 5 incorrect project progress summaries can be generated, namely: Incorrect Project Progress Summary 1 (clause A, clause C, clause B), Incorrect Project Progress Summary 2 (clause B, clause A, clause C), Incorrect Project Progress Summary 3 (clause B, clause C, clause A), Incorrect Project Progress Summary 4 (clause C, clause A, clause B), and Incorrect Project Progress Summary 5 (clause C, clause B, clause A).

[0112] s33. Using the correct project schedule summary and K incorrect project schedule summaries, construct K sequence sample pairs. Each sequence sample pair includes a positive sample and a negative sample. The positive sample in any sequence sample pair is the correct project schedule summary, and the negative sample in any sequence sample pair is an incorrect project schedule summary. The negative samples in different sequence sample pairs are different.

[0113] Step s34: Obtain the positive and negative example generation probabilities for each sequence sample pair, and calculate the ranking loss value of the language model based on these probabilities. Specifically, the loss value for each sequence sample pair can be calculated based on its positive and negative example generation probabilities, using the same method as the method described above for calculating the loss value of the original sample pairs, which will not be repeated here. Furthermore, the loss values ​​of each sequence sample pair can be integrated to obtain the ranking loss value of the language model; a schematic diagram illustrating the calculation logic of this ranking loss value using 5 sequence sample pairs as an example can be found in [link to example]. Figure 5d As shown.

[0114] Alternatively, a specific implementation for calculating the ranking loss value of the language model based on the positive and negative example generation probabilities of each sequence sample pair can be: obtaining the ranking quality deviation score of the negative sample in each sequence sample pair. This ranking quality deviation score indicates the degree of deviation between the ranking quality of the negative sample and the ranking quality of the corresponding positive sample. Specifically, the ranking quality score of the positive sample and the ranking quality score of the negative sample in each sequence sample pair can be calculated separately, thereby calculating the ratio between the ranking quality score of each negative sample and the ranking quality score of the positive sample, which serves as the ranking quality deviation score of the corresponding negative sample. As described above, the negative sample in each sequence sample pair is obtained by rearranging the clauses in the positive sample. That is, each negative sample and each positive sample includes I clauses, differing only in their order. Therefore, a relevance score can be defined for each clause based on its position in the positive sample. The DCG formula can then be used to calculate the ranking quality score of the corresponding sample based on the relevance score of each clause and its position in any sample (positive or negative). The DCG formula is shown in Formula 3.1 below.

[0115]

[0116] In Formula 3.1 above, i represents the sequence number of the clause, rel i This represents the relevance score of the clause with sequence number i. When calculating the ranking quality score of positive and negative samples using Formula 3.1 above, the formula for calculating the ranking quality deviation score of any negative sample can be found in Formula 3.2 below:

[0117]

[0118] In Formula 3.2 above, IDCG represents the ranking quality score of the positive sample, DCG represents the ranking quality score of any negative sample, and nDCG represents the ranking quality deviation score of the corresponding negative sample.

[0119] After obtaining the ranking quality deviation score of the negative samples in each sequence sample pair, the ranking loss value of the language model can be calculated based on the ranking quality deviation score of the negative samples in each sequence sample pair, as well as the positive and negative example generation probabilities of each sequence sample pair. It is understandable that, since the ranking quality deviation score of the negative samples in a ranked sample pair is equal to the ratio between the ranking quality score of the negative samples and the ranking quality score of the positive samples, the negative example generation probability (π) corresponding to the negative samples in the ranked sample pair should be: θ (y L |x)) / Probability of generating a positive example corresponding to the positive example sample (π) θ (yw |x))≤ the ranking quality deviation score of the corresponding negative sample, that is, π should be present. θ (y L |x) / π θ (y w |x)≤nDCG (the probability of negative samples (error project progress summary) should be as small as possible, so it is less than or equal to)), therefore we should have: logπ θ (y w |x)-logπ θ (y L |x)≥log(1 / nDCG). Based on this, the specific formula for calculating the ranking loss value of the language model, based on the ranking quality deviation score of the negative samples in each sequence sample pair, and the positive and negative generation probabilities of each sequence sample pair, can be found in Formula 3.3 below:

[0120]

[0121] In Formula 3.3 above, Loss3 represents the ranking loss value, G represents the set containing all sequence sample pairs, B = log(1 / nDCG), and the definitions of other parameters can be found in the relevant description of Formula 2.1 above, which will not be repeated here.

[0122] It should be noted that steps s31-s34 above are merely illustrative descriptions of one implementation method for a computer device to calculate the ranking loss value of a language model, and are not intended to limit the scope of the implementation. For example, in other embodiments, the computer device may also construct multiple ranked sample pairs based on steps s31-s33 above, calculate the sample difference between positive and negative samples in each ranked sample pair, obtain the loss value of each ranked sample pair, and then integrate the loss values ​​of each ranked sample pair to obtain the ranking loss value of the language model.

[0123] S406 optimizes the model parameters of a language model based on the model loss value.

[0124] When the model loss value includes only one of the correctness loss value, expression loss value, and ranking loss value, the specific implementation of step S406 can be: optimizing the model parameters of the language model in the direction of reducing the model loss value. Here, "in the direction of reducing the model loss value" means: optimizing the model in a direction that aims to minimize the model loss value; optimizing the model in this direction ensures that the model loss value generated by the language model after each optimization is less than the model loss value generated by the language model before optimization. For example, if the calculated model loss value is 0.85, then after optimizing the model parameters of the language model in the direction of reducing the model loss value, the model loss value generated by the optimized language model should be less than 0.85.

[0125] When the model loss value includes correctness loss, expression loss, and ranking loss, step S406 can be implemented as follows: The mean of each loss value in the model loss value is calculated to obtain the target loss value, and the model parameters of the language model are optimized based on the target loss value. Alternatively, the weights corresponding to each loss value in the model loss value are obtained, where each loss value includes correctness loss, expression loss, and ranking loss; the obtained weights are used to perform a weighted sum of each loss value to obtain the target loss value, and the model parameters of the language model are optimized based on the target loss value, such as optimizing the model parameters of the language model in the direction of reducing the target loss value. For example, if a1 represents the weight of the correctness loss value, a2 represents the weight of the expression loss value, and a3 represents the weight of the ranking loss value, then the target loss value Loss = a1*Loss1 + a2*loss2 + a3*loss3.

[0126] The weights of the aforementioned loss values ​​can be set based on actual needs, and this application embodiment does not limit the magnitude relationship between the weights of the various loss values. In an optional embodiment, the weight corresponding to the correctness loss value can be greater than the weight corresponding to the expression loss value, and the weight corresponding to the expression loss value can be greater than the weight corresponding to the ranking loss value. For example, the weight a1 corresponding to the correctness loss value can be equal to 0.7 (i.e., a1 = 0.7), the weight a2 corresponding to the expression loss value can be 0.25 (i.e., a2 = 0.25), and the weight corresponding to the ranking loss value can be 0.05 (i.e., a3 = 0.05). This allows the language model to be optimized based on these three loss values, with these three loss values ​​acting on the language model in a progressive manner. This enables the language model to progressively improve its ability to generate correct project progress summaries, express reasonableness, and rank, thereby improving the overall performance of the language model.

[0127] It should be noted that when optimizing the parameters of a language model, one can optimize all or only some of the model parameters; there is no limitation on this. For example, in a language model comprising 64 decoder layers, normalization layers, and an lm_head structure, the model parameters involved in the process of generating the input text embedding through the normalization layers and mapping the input text embedding to the dictionary vector space through the lm_head structure during language model training do not need to be included in the training optimization. The ModuleList (model parameter list) that needs to be optimized can include the model parameters in the 64 decoder layers. This reduces the number of model parameters to be optimized during training, thereby improving model training efficiency. For example, the pseudocode for the training logic of the language model in this case can be as follows:

[0128]

[0129] Based on the above description, the general logic of the information processing method for training a language model proposed in the embodiments of this application can be summarized as follows: Figure 5e As shown, it can have at least the following advantages:

[0130] 1. Capable of integrating and processing data from various sources and personnel: It organizes data from different sources and in different formats (such as tables and text descriptions) into progress records and records the entry time. Subsequently, the language model directly integrates this information to generate a project progress summary. The entire process can be automated, reducing the consumption of application resources.

[0131] 2. Achieve domain knowledge embedding: By training the language model to learn the rules of summarizing and expressing in the media production domain (such as the requirements for submission and the rules of language wording style), the language model can reduce the fuzzy or inappropriate output of the language model under domain professional summary tasks, improve the accuracy of the project progress summary generated by the language model, and reduce human intervention, thereby reducing labor costs.

[0132] 3. Optimize Output Time Order: By optimizing the language model's parameters, the order of clauses in the project progress summary output by the language model can be ensured to be correct. This allows the language model to accurately output the corresponding project progress summary according to the project organization (i.e., media production process) even when faced with various common out-of-order input progress records. This makes the timeline of the project progress summary clear at a glance, improving the accuracy of the project progress summary. Furthermore, when the language model detects uncertain clauses in the project progress summary (such as "contradictory submission times"), it can issue an early warning, making it easier for relevant personnel to perceive these uncertain clauses and facilitating manual intervention.

[0133] Based on the above description, this application also proposes a method for summarizing the progress of media projects (such as film and television production projects). This method can train a self-developed language model so that the trained language model can automatically sort out relevant knowledge in the field of media production (such as film and television production) and summarize the progress record information uploaded by multiple parties. Based on multiple progress record information and timelines, it can generate corresponding project progress summaries and information displays, and give corresponding warnings when the project progress summary contains clauses with uncertainty.

[0134] One training logic for a language model can include the following parts:

[0135] 1. Data Acquisition: Specifically, multiple progress records from various sample media projects can be collected. Input text x is constructed using these progress records for each sample media project. A language model is then called to summarize each input text x, resulting in multiple initial project progress summaries yr0. This generates multiple samples (x, yr0). The process of generating the initial project progress summaries yr0 is the model forward pass, roughly as follows: Input information x is used as model input (inputtxt). The first predicted character id1 of the language model is obtained. id1 is concatenated with historical input txt and input into the language model to obtain id2, and so on, until n_i outputs are obtained (n_i is the number of characters in the output items of the sample). This yields the initial project progress summary yr0 and the prediction probability matrix (n_i*152064, representing the prediction probability of each character in the dictionary).

[0136] 2. Data Construction: For each sample (x, yr0), generate 2 or 4 original sample pairs and K sequence sample pairs as training data, along with parameters related to the training data, as detailed below:

[0137] (1) Annotation optimization generates multiple original sample pairs. Specifically, for the initial project progress summary yr0, manual verification and summarization can be performed. First, ensure that the descriptions of each clause in the initial project progress summary yr0 meet the requirements and avoid ambiguity, resulting in the intermediate project progress summary yr01. Then, arrange the clauses in the intermediate project progress summary yr01 in chronological order and ensure that the logical relationship between the execution progress described by each clause conforms to the project development logic (i.e., the logical relationship specified by the media production process), resulting in the target project progress summary yr02. This generates two original sample pairs (x, yr02, yr0) and (x, yr02, yr01). Among them, yr02 represents a good sample, and the others represent bad samples (yr: y raw means the original sample output). If the language model is only called to summarize the input text x once, the target progress summary yr02 can be used as the subsequent correct project progress summary y0.

[0138] Optionally, the language model can be invoked to perform a second summary of x, resulting in an initial project progress summary yr1. This initial project progress summary yr1 is then manually validated to generate an intermediate project progress summary yr11 and a target project progress summary yr12, thus generating two more original sample pairs: (x, yr12, yr1) and (x, yr12, yr11). Therefore, in this case, a total of four original sample pairs can be generated for the sample (x, yr0). The target project progress summaries yr02 and yr12 can be compared to select one as the correct subsequent project progress summary y0. It is understandable that ambiguity and order issues often recur in multiple predictions, so multiple predictions can capture more relevant errors. For ease of explanation, the following explanation will use the example of summarizing the input text x only once, i.e., without performing a second summary of the input text x.

[0139] (2) Generate K sequence sample pairs highlighting sequence information. Specifically, for the selected correct project progress summary y0, its clauses have a time series dependency relationship. Therefore, comparison sample pairs can be generated for the preceding and following clauses. That is, for each input text x and the corresponding correct project progress summary y0, the clauses in the correct project progress summary y0 can be rearranged to generate K incorrect project progress summaries, thereby generating K sequence sample pairs. Taking the correct project progress summary y0 as having 3 clauses and each clause representing a piece of time information as an example, the logic for generating K sequence sample pairs can be as follows:

[0140] Input: x

[0141] Correct project progress summary: y0(Time information 1, Time information 2, Time information 3)

[0142] Summary of the progress of K erroneous projects: y1(Time information 1, Time information 3, Time information 2), y2(Time information 2, Time information 1, Time information 3), y3(Time information 2, Time information 3, Time information 1), y4(Time information 3, Time information 1, Time information 2), y5(Time information 3, Time information 2, Time information 1).

[0143] There are K sequence sample pairs: (x, y0, y1), (x, y0, y2)……(x, y0, y5), a total of 5 sequence sample pairs, among which y1~y5 are relatively worse.

[0144] (3) Calculation of weights for clauses in sequence samples:

[0145] For all clauses in y0 (the correct answer) above (one time information x is one clause), the relevance scores (rel) of "time information 1, time information 2, and time information 3" are defined as 3, 2, and 1 respectively (for the case where y0 includes 4 clauses (4 time information), the relevance scores of each clause are 4, 3, 2, and 1 respectively, and so on).

[0146] Calculate the score rel for each clause (time information). i / log2(i+1), recorded as s1, s2, s3 in sequence, so as to assign weights to each clause (time information) and obtain the following results: the weight of time information 1 is s1, the weight of time information 2 is s2, and the weight of time information 3 is s3 - these weights can be used for subsequent weighting of y0 by Loss1.

[0147] For y1, calculate its DCG score according to the DCG formula (i is the order number of the clause) and calculate the DCG score of y0 as IDCG, thereby calculating the nDCG score of y1. For y2 to y5, calculate their corresponding nDCG scores in the same way. These nDCG scores can be used for the subsequent weighting of different sequence sample pairs in Loss3.

[0148] 3. Training, specifically including the following:

[0149] 1) Parameter initialization: The network weights of the open-source pre-trained base model are used to initialize the model parameters of the language model.

[0150] 2) Set learning parameters: Use LoRA (Low Rank Adaptation) technique for fine-tuning, and only train the projection matrices of q (query), k (key), and v (value) of the attention structure in the decoder of the language model.

[0151] 3) Learning process: Each set of bs sample pairs is used as a batch for training. The entire set of samples is traversed once to complete one epoch. The bs sample pairs consist of 2+K sample pairs from the same sample (x, yr0) (i.e., 2 original sample pairs and K sequence sample pairs).

[0152] In the process of learning from the full dataset in one round, during each batch of learning, b / s sample pairs can be used to calculate the model loss value. Based on this loss value, the model backward propagation process is executed, which involves backpropagating the parameter gradients of the corresponding model parameters in the language model based on the loss value, and then updating the corresponding model parameters according to the calculated gradients. The specific method for calculating the model loss value using b / s sample pairs can be as follows:

[0153] ① Use the correct project progress summary y0 to calculate Loss1. Loss1 can be the cross-entropy loss trained on the model based on nDCG gain.

[0154] ② Loss2 is calculated using two original sample pairs. Specifically, the loss value of each original sample pair (x, yr02, yr0) and (x, yr02, yr01) can be calculated separately, and the average of the loss values ​​of the two original sample pairs can be calculated as Loss2. Since the sample loss is related to ambiguity and order, this loss value Loss2 belongs to the expression loss.

[0155] ③ Calculate Loss3 using K sequence sample pairs. Specifically, calculate the loss value of each sequence sample in the K sequence sample pairs, and then calculate the average of the loss values ​​of these K sequence sample pairs as Loss3. In other words, calculate the loss value for all sequential sequences of the correct project progress summary y0 and the positive and negative answers constructed by y0 (i.e., sequence sample pairs), and then calculate the average of the loss values ​​as Loss3. This Loss3 belongs to the sequence sample pair loss.

[0156] ④ The model loss values ​​are obtained by constructing Loss1, Loss2 and Loss3.

[0157] It should be noted that after training the language model using the above logic, the trained language model can be applied to the progress management scenario of film and television production projects. The application process can be found in [reference needed]. Figure 6a As shown, it generally includes the following steps:

[0158] (1) Input records: Specifically, the progress record information of the film and television production project can be input by the personnel of all parties involved in the film and television production project. This progress record information can be input through specific media on terminals such as mobile phones and computers (such as voice input, which is converted into text by the system, so as to facilitate the input of progress record information by personnel with different operating habits). The progress record information entered by any person shall include at least: time information, project identifier, and stage progress information (used to indicate the execution progress of the project stage); for example, taking the various parties involved in a film and television production project, including screenwriters, producers, reviewers, post-production personnel such as editors, and operations personnel, the progress record information entered by screenwriters may be, for example, "Project 2 2022 / 4 / 20 script adaptation completed" and "Project 2 2022 / 10 / 13 script review completed", the progress record information entered by producers may be, for example, "Project 2 2023 / 1 / 5 filming started", "Project 2 2023 / 8 / 5 two-thirds of filming completed" and "Project 2 2024 / 1 / 12 36 episodes completed", the progress record information entered by reviewers may be, for example, "Project 2 2025 / 3 / 20 submitted to review agency A, awaiting feedback from review agency A" and "Project 2 "2025 / 5 / 11 Review agency A approved, waiting for the director's final version to be submitted to review agency B." The progress record information entered by the editing and other post-production staff can be, for example, "Project 2 2025 / 2 / 18 Director's cut completed, music released." The progress record information entered by the operations staff can be, for example, "Project 2 2022 / 11 / 13 Script adaptation quality excellent, payment made."

[0159] (2) Archive by project and time. Specifically, the progress record information input by all parties can be archived according to project (category) and time. For example, from the progress record information input by all parties, each progress record containing the same project identifier can be filtered. According to the stage progress information in each progress record, the project stage corresponding to each progress record can be determined. The progress record information corresponding to the same project stage can be arranged in chronological order from earliest to latest to obtain the archive result. Optionally, the archive result can be displayed in the interface in a timeline style, such as... Figure 6b As shown; alternatively, based on the archived results, multiple time nodes and corresponding execution progress for each film and television production project can be compiled and displayed on the interface. Special events and abnormal progress (abnormal execution progress) can also be highlighted (e.g., highlighted with a specified color). Therefore, this step can merge progress records from multiple sources at the same or different time nodes into the corresponding project stage of the same film and television production project, thus eliminating the problem of incomplete summaries caused by fragmented information.

[0160] (3) Model summary: Specifically, the trained model can be called to perform semantic understanding of the progress record information corresponding to each project stage in the archived results, summarize the project stages, task nodes, progress status and abnormal events, and generate a project progress summary based on the summary results. For example, the generated project progress summary can be "...passed the review of auditing agency A on May 11, 2025, and is scheduled to be submitted to auditing agency B for review on May 30, 2025...".

[0161] (4) Output project progress summary. Specifically, the project progress summary can be automatically pushed to relevant personnel (such as relevant responsible persons and management). Alternatively, it can be displayed as a timeline on the terminal interface of relevant responsible persons and management, or as a summary display page. Figure 6c As shown. Optionally, it can also support users to customize subscription nodes. That is, with the support of various parties, users can set the project stages they are interested in according to their needs. If the project stage summarized by the language model is the project stage that the user is interested in, the project progress summary will be pushed to the user.

[0162] In summary, the embodiments of this application can train a language model using comparative samples constrained by chronological order, solving the problem of disorganized timelines in imported information and reducing the ambiguity of output descriptions due to information overload in project progress summaries. Furthermore, it can automatically aggregate fragmented information from multiple sources into unified project progress information, process it using model methods to obtain more accurate project progress summaries with a more chronological order, and provide risk perception and early warnings to facilitate operations personnel in tracking project progress, improving project management efficiency. The summary (and early warning) time is reduced from tens of hours of manual work to minutes.

[0163] Based on the descriptions of the various method embodiments above, this application also discloses an information processing apparatus; this information processing apparatus can execute each step in any of the above method flows. Please refer to... Figure 7 The information processing device can operate the following units:

[0164] The acquisition unit 701 is used to acquire multiple progress record information of the sample media project. The sample media project is executed according to the media production process. One progress record information records the execution progress of the sample media project under a certain time information. The media production process defines the logical relationship between different execution progress.

[0165] Processing unit 702 is used to call the language model to summarize based on multiple progress record information of the sample media project to obtain an initial project progress summary. The initial project progress summary includes multiple clauses arranged in sequence, and one clause is used to describe the execution progress of the sample media project under a certain time information.

[0166] The processing unit 702 is further configured to calibrate multiple clauses in the initial project progress summary based on the time information and execution progress described in each clause of the initial project progress summary to obtain a target project progress summary; the multiple clauses in the target project progress summary are arranged in chronological order, and the logical relationship between the execution progress described in the multiple clauses matches the logical relationship specified in the media production process.

[0167] The processing unit 702 is further configured to train the language model using the target project progress summary, and the trained language model is used to generate a corresponding media project progress summary based on multiple progress record information of any media project.

[0168] In one specific implementation, when calibrating multiple clauses in the initial project progress summary to obtain the target project progress summary, the processing unit 702 may be specifically used for:

[0169] Based on the summary and expression guidelines in the field of media production, the expression of each clause in the initial project progress summary is calibrated to obtain the intermediate project progress summary.

[0170] In the intermediate project progress summary, the clauses are rearranged according to the time information and execution progress described in each clause to obtain the target project progress summary.

[0171] In another specific implementation, when training the language model using the target project progress summary, the processing unit 702 may be specifically used for:

[0172] Based on the progress summary of the target project, the model loss value of the language model is calculated. The model loss value includes at least one of the following: correctness loss value, expression loss value, and ranking loss value. The correctness loss value refers to the loss value generated by the language model in the dimension of summary correctness, the expression loss value refers to the loss value generated by the language model in the dimension of summary expression rationality, and the ranking loss value refers to the loss value generated by the language model in the dimension of clause ranking.

[0173] Based on the model loss value, the model parameters of the language model are optimized.

[0174] In another specific implementation, the model loss value includes a correctness loss value, and the number of target project progress summaries is N, where N is a positive integer;

[0175] Accordingly, when calculating the model loss value of the language model based on the target project progress summary, the processing unit 702 can be specifically used for:

[0176] From the N target project progress summaries, select one target project progress summary as the correct project progress summary;

[0177] Each clause in the correct project progress summary is weighted and scored to obtain a weight score for each clause in the correct project progress summary. The weight score is negatively correlated with the position of the corresponding clause in the correct project progress summary.

[0178] Obtain the generation probability of each character in the correct project progress summary, where the generation probability refers to the probability that the language model generates the corresponding character;

[0179] Based on the generation probability of each character in the correct project progress summary and the weight score of each clause, the correctness loss value of the language model is calculated.

[0180] In another specific implementation, the model loss value includes the expression loss value, the target project progress summary is obtained by calibrating the order of the clauses in the intermediate project progress summary, and the intermediate project progress summary is obtained by calibrating the expression of the clauses in the initial project progress summary.

[0181] Accordingly, when calculating the model loss value of the language model based on the target project progress summary, the processing unit 702 can be specifically used for:

[0182] Multiple original sample pairs are constructed using the target project progress summary, the intermediate project progress summary, and the initial project progress summary; wherein each original sample pair includes positive sample and negative sample, the positive sample in any original sample pair is the target project progress summary, and the negative sample in any original sample pair is either the intermediate project progress summary or the initial project progress summary.

[0183] Obtain the positive example generation probability and negative example generation probability for each original sample pair. The positive example generation probability refers to the probability that the language model generates a positive example sample, and the negative example generation probability refers to the probability that the language model generates a negative example sample.

[0184] The expression loss value of the language model is calculated based on the positive and negative example generation probabilities of each original sample pair.

[0185] In another specific implementation, the model loss value includes a ranking loss value, and the number of target project progress summaries is N, where N is a positive integer;

[0186] Accordingly, when calculating the model loss value of the language model based on the target project progress summary, the processing unit 702 can be specifically used for:

[0187] From the N target project progress summaries, select one target project progress summary as the correct project progress summary;

[0188] The clauses in the correct project progress summary are rearranged K times to obtain K incorrect project progress summaries. Each rearrangement is used to generate one incorrect project progress summary, where K is a positive integer.

[0189] Using the correct project progress summary and the K incorrect project progress summaries, construct K sequence sample pairs; wherein, each sequence sample pair includes a positive sample and a negative sample, the positive sample in any sequence sample pair is the correct project progress summary, and the negative sample in any sequence sample pair is an incorrect project progress summary;

[0190] Obtain the positive example generation probability and negative example generation probability of each sequence sample pair, and calculate the ranking loss value of the language model based on the positive example generation probability and negative example generation probability of each sequence sample pair.

[0191] In another specific implementation, when calculating the ranking loss value of the language model based on the positive example generation probability and negative example generation probability of each sequence sample pair, the processing unit 702 may be specifically used for:

[0192] Obtain the ranking quality deviation score of the negative sample in each sequence sample pair. The ranking quality deviation score of the negative sample is used to indicate the degree of deviation between the ranking quality of the negative sample and the ranking quality of the corresponding positive sample.

[0193] The ranking loss value of the language model is calculated based on the ranking quality deviation score of the negative sample in each sequence sample pair, and the positive and negative sample generation probabilities of each sequence sample pair.

[0194] In another specific implementation, when the model loss value includes a correctness loss value, an expression loss value, and a ranking loss value, when optimizing the model parameters of the language model based on the model loss value, the processing unit 702 may specifically be used to:

[0195] Obtain the weights corresponding to each loss value in the model loss values, wherein the weight corresponding to the correctness loss value is greater than the weight corresponding to the expression loss value, and the weight corresponding to the expression loss value is greater than the weight corresponding to the ranking loss value;

[0196] The obtained weights are used to sum the various loss values ​​to obtain the target loss value, and the model parameters of the language model are optimized based on the target loss value.

[0197] In another specific implementation, after obtaining the initial project progress summary, the processing unit 702 may be specifically used for:

[0198] The language model is invoked to perform uncertainty detection on each clause in the initial project progress summary, and the detection results are obtained; wherein, the uncertainty of the clause is used to indicate the credibility of the language model for the corresponding clause;

[0199] If the detection result indicates that there is an uncertain clause in the initial project progress summary, then an early warning will be issued for the initial project progress summary.

[0200] According to another embodiment of this application, Figure 7 The various units in the information processing apparatus shown can be individually or entirely merged into one or more other units, or some of the units can be further divided into multiple functionally smaller units. This achieves the same operation without affecting the technical effects of the embodiments of this application. The above-mentioned units are divided based on logical functions. In practical applications, the function of one unit can also be implemented by multiple units, or the function of multiple units can be implemented by one unit. In other embodiments of this application, the information processing apparatus may also include other units. In practical applications, these functions can also be implemented with the assistance of other units, and can be implemented collaboratively by multiple units.

[0201] According to another embodiment of this application, a computer program (including one or more instructions) capable of performing the steps involved in any of the above methods can be run on a general-purpose computing device, such as a computer, which includes processing elements and storage elements such as a central processing unit (CPU), random access memory (RAM), and read-only memory (ROM), to construct a system such as... Figure 7 The information processing apparatus shown herein, and the methods implemented in the embodiments of this application. The computer program may be recorded on, for example, a computer-readable storage medium, loaded onto the aforementioned computing device via the computer-readable storage medium, and run therein.

[0202] It is worth noting that, in the embodiments of this application, the terms "module" or "unit" refer to a computer program or part of a computer program with a predetermined function, which works together with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can contain a portion of the overall module or unit's functionality.

[0203] This application embodiment can acquire multiple progress record information of sample media projects executed according to the media production process, and call a language model to generate an initial project progress summary based on these multiple progress record information. Based on the time information and execution progress described in each clause of the initial project progress summary, the multiple clauses in the initial project progress summary are calibrated to obtain a target project progress summary. This ensures that the multiple clauses in the target project progress summary are arranged in chronological order, and that the logical relationships between the execution progress described in the multiple clauses match the logical relationships specified in the media production process. This guarantees that the target project progress summary has correct temporal and logical relationships. Therefore, when using the target project progress summary to train the language model, the language model can better understand the logical relationships between different execution progresses and improve its time series summarization ability during training. This enhances the logic and temporal rationality of the language model when generating project progress summaries. As a result, when the trained language model generates project progress summaries based on multiple progress record information of any media project, the risk of temporal confusion and logical errors in the project progress summary can be reduced, thereby improving the accuracy of the project progress summary. Furthermore, since the embodiments of this application automatically generate project progress summaries for any media project by training a language model and based on the trained language model, the embodiments of this application can realize the entire process of generating project progress summaries without the participation of professional personnel. This not only saves labor costs, reduces the risk of human error, and improves the accuracy of project progress summaries, but also avoids the problem of low efficiency in generating project progress summaries due to low human efficiency, thereby improving the efficiency of generating project progress summaries.

[0204] Based on the description of the above method and apparatus embodiments, this application also provides a computer device. Please refer to... Figure 8The computer device includes at least a processor 801, an input interface 802, an output interface 803, and a computer storage medium 804. The processor 801, input interface 802, output interface 803, and computer storage medium 804 within the computer device can be connected via a bus or other means. The computer storage medium 804 can be stored in the computer device's memory. The computer storage medium 804 is used to store a computer program, which includes one or more instructions. The processor 801 is used to execute one or more instructions from the computer program stored in the computer storage medium 804. The processor 801 (or CPU (Central Processing Unit)) is the computing and control core of the computer device, adapted to implement one or more instructions, specifically adapted to load and execute one or more instructions to achieve a corresponding method flow or corresponding function. In one embodiment, the processor 801 described in this application can be used to perform corresponding steps in any of the above-mentioned information processing methods.

[0205] This application also provides a computer storage medium (Memory), which is a memory device in a computer device used to store computer programs and data. It is understood that the computer storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer storage medium provides storage space that stores the operating system of the computer device. Furthermore, the storage space also stores a computer program, which includes one or more instructions suitable for loading and execution by a processor 801. These instructions can be one or more program codes. It should be noted that the computer storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device; optionally, it can also be at least one computer storage medium located remotely from the aforementioned processor. In one embodiment, the processor can load and execute one or more instructions stored in the computer storage medium to implement the corresponding steps in any of the above method embodiments.

[0206] This application embodiment can acquire multiple progress record information of sample media projects executed according to the media production process, and call a language model to generate an initial project progress summary based on these multiple progress record information. Based on the time information and execution progress described in each clause of the initial project progress summary, the multiple clauses in the initial project progress summary are calibrated to obtain a target project progress summary. This ensures that the multiple clauses in the target project progress summary are arranged in chronological order, and that the logical relationships between the execution progress described in the multiple clauses match the logical relationships specified in the media production process. This guarantees that the target project progress summary has correct temporal and logical relationships. Therefore, when using the target project progress summary to train the language model, the language model can better understand the logical relationships between different execution progresses and improve its time series summarization ability during training. This enhances the logic and temporal rationality of the language model when generating project progress summaries. As a result, when the trained language model generates project progress summaries based on multiple progress record information of any media project, the risk of temporal confusion and logical errors in the project progress summary can be reduced, thereby improving the accuracy of the project progress summary. Furthermore, since the embodiments of this application automatically generate project progress summaries for any media project by training a language model and based on the trained language model, the embodiments of this application can realize the entire process of generating project progress summaries without the participation of professional personnel. This not only saves labor costs, reduces the risk of human error, and improves the accuracy of project progress summaries, but also avoids the problem of low efficiency in generating project progress summaries due to low human efficiency, thereby improving the efficiency of generating project progress summaries.

[0207] It should be noted that, according to one aspect of this application, a computer program product or computer program is also provided, comprising one or more instructions stored in a computer storage medium. A processor of a computer device reads one or more instructions from the computer storage medium and executes the one or more instructions, causing the computer device to perform the methods provided in various optional embodiments of the above-described methods. It should be understood that the above-disclosed embodiments are merely preferred embodiments of this application and should not be construed as limiting the scope of this application. Therefore, equivalent variations made according to the claims of this application are still within the scope of this application.

Claims

1. An information processing method, characterized in that, include: Multiple progress records of a sample media project are obtained. The sample media project is executed according to the media production process. Each progress record records the execution progress of the sample media project at a certain time. The media production process defines the logical relationship between different execution progresses. The language model is invoked to summarize the progress information of the sample media project based on multiple progress records to obtain an initial project progress summary. The initial project progress summary includes multiple clauses arranged in sequence, and each clause is used to describe the execution progress of the sample media project under a certain time information. Based on the time information and execution progress described in each clause of the initial project progress summary, multiple clauses in the initial project progress summary are calibrated to obtain the target project progress summary; the multiple clauses in the target project progress summary are arranged in chronological order, and the logical relationship between the execution progress described in the multiple clauses matches the logical relationship specified in the media production process; The language model is trained using the target project progress summary. The trained language model is then used to generate a corresponding media project progress summary based on multiple progress record information of any media project.

2. The method as described in claim 1, characterized in that, The calibration of multiple clauses in the initial project progress summary to obtain the target project progress summary includes: Based on the summary and expression guidelines in the field of media production, the expression of each clause in the initial project progress summary is calibrated to obtain the intermediate project progress summary. In the intermediate project progress summary, the clauses are rearranged according to the time information and execution progress described in each clause to obtain the target project progress summary.

3. The method as described in claim 1 or 2, characterized in that, The step of using the target project progress summary to train the language model includes: Based on the progress summary of the target project, the model loss value of the language model is calculated. The model loss value includes at least one of the following: correctness loss value, expression loss value, and ranking loss value. The correctness loss value refers to the loss value generated by the language model in the dimension of summary correctness, the expression loss value refers to the loss value generated by the language model in the dimension of summary expression rationality, and the ranking loss value refers to the loss value generated by the language model in the dimension of clause ranking. Based on the model loss value, the model parameters of the language model are optimized.

4. The method as described in claim 3, characterized in that, The model loss value includes the correctness loss value, and the number of target project progress summaries is N, where N is a positive integer; The calculation of the model loss value of the language model based on the progress summary of the target project includes: From the N target project progress summaries, select one target project progress summary as the correct project progress summary; Each clause in the correct project progress summary is weighted and scored to obtain a weight score for each clause in the correct project progress summary. The weight score is negatively correlated with the position of the corresponding clause in the correct project progress summary. Obtain the generation probability of each character in the correct project progress summary, where the generation probability refers to the probability that the language model generates the corresponding character; Based on the generation probability of each character in the correct project progress summary and the weight score of each clause, the correctness loss value of the language model is calculated.

5. The method as described in claim 3, characterized in that, The model loss value includes the expression loss value; the target project progress summary is obtained by calibrating the order of the clauses in the intermediate project progress summary; and the intermediate project progress summary is obtained by calibrating the expression of the clauses in the initial project progress summary. The calculation of the model loss value of the language model based on the progress summary of the target project includes: Multiple original sample pairs are constructed using the target project progress summary, the intermediate project progress summary, and the initial project progress summary; wherein each original sample pair includes positive sample and negative sample, the positive sample in any original sample pair is the target project progress summary, and the negative sample in any original sample pair is either the intermediate project progress summary or the initial project progress summary. Obtain the positive example generation probability and negative example generation probability for each original sample pair. The positive example generation probability refers to the probability that the language model generates a positive example sample, and the negative example generation probability refers to the probability that the language model generates a negative example sample. The expression loss value of the language model is calculated based on the positive and negative example generation probabilities of each original sample pair.

6. The method as described in claim 5, characterized in that, The language model performs N summaries based on multiple progress record information of the sample media project to obtain N initial project progress summaries, where N is a positive integer. When calibrating multiple clauses in the nth initial project progress summary, we obtain the nth intermediate project progress summary and the nth target project progress summary, where n∈[1,N]; The progress summary of the nth target project and the progress summary of the nth initial project form an original sample pair, and the progress summary of the nth target project and the progress summary of the nth intermediate project form an original sample pair.

7. The method as described in claim 3, characterized in that, The model loss value includes the ranking loss value, and the number of target project progress summaries is N, where N is a positive integer. The calculation of the model loss value of the language model based on the progress summary of the target project includes: From the N target project progress summaries, select one target project progress summary as the correct project progress summary; The clauses in the correct project progress summary are rearranged K times to obtain K incorrect project progress summaries. Each rearrangement is used to generate one incorrect project progress summary, where K is a positive integer. Using the correct project progress summary and the K incorrect project progress summaries, construct K sequence sample pairs; wherein, each sequence sample pair includes a positive sample and a negative sample, the positive sample in any sequence sample pair is the correct project progress summary, and the negative sample in any sequence sample pair is an incorrect project progress summary; Obtain the positive example generation probability and negative example generation probability of each sequence sample pair, and calculate the ranking loss value of the language model based on the positive example generation probability and negative example generation probability of each sequence sample pair.

8. The method as described in claim 7, characterized in that, The calculation of the ranking loss value of the language model based on the positive example generation probability and negative example generation probability of each sequence sample pair includes: Obtain the ranking quality deviation score of the negative sample in each sequence sample pair. The ranking quality deviation score of the negative sample is used to indicate the degree of deviation between the ranking quality of the negative sample and the ranking quality of the corresponding positive sample. The ranking loss value of the language model is calculated based on the ranking quality deviation score of the negative sample in each sequence sample pair, and the positive and negative sample generation probabilities of each sequence sample pair.

9. The method as described in claim 3, characterized in that, When the model loss value includes correctness loss, expression loss, and ranking loss, optimizing the model parameters of the language model based on the model loss value includes: Obtain the weights corresponding to each loss value in the model loss values, wherein the weight corresponding to the correctness loss value is greater than the weight corresponding to the expression loss value, and the weight corresponding to the expression loss value is greater than the weight corresponding to the ranking loss value; The obtained weights are used to sum the various loss values ​​to obtain the target loss value, and the model parameters of the language model are optimized based on the target loss value.

10. The method according to any one of claims 1, 2, or 4-9, characterized in that, The media production process includes multiple project phases, and each progress record includes both fixed and variable description fields. The fixed description field includes time information and stage progress information, and the variable description field includes remarks related to the stage progress information. The stage progress information is used to indicate the execution progress of the sample media project at the corresponding time stage.

11. The method according to any one of claims 1, 2, or 4-9, characterized in that, After obtaining the initial project progress summary, the method further includes: The language model is invoked to perform uncertainty detection on each clause in the initial project progress summary, and the detection results are obtained; wherein, the uncertainty of the clause is used to indicate the credibility of the language model for the corresponding clause; If the detection result indicates that there is an uncertain clause in the initial project progress summary, then an early warning will be issued for the initial project progress summary.

12. An information processing device, characterized in that, include: The acquisition unit is used to acquire multiple progress record information of the sample media project. The sample media project is executed according to the media production process. One progress record information records the execution progress of the sample media project under a certain time information. The media production process defines the logical relationship between different execution progress. The processing unit is used to call the language model to summarize based on multiple progress record information of the sample media project to obtain an initial project progress summary. The initial project progress summary includes multiple clauses arranged in sequence, and one clause is used to describe the execution progress of the sample media project under a certain time information. The processing unit is further configured to calibrate multiple clauses in the initial project progress summary based on the time information and execution progress described in each clause of the initial project progress summary to obtain a target project progress summary; the multiple clauses in the target project progress summary are arranged in chronological order, and the logical relationship between the execution progress described in the multiple clauses matches the logical relationship specified in the media production process; The processing unit is further configured to train the language model using the target project progress summary, and the trained language model is used to generate a corresponding media project progress summary based on multiple progress record information of any media project.

13. A computer device, comprising an input interface and an output interface, characterized in that, Also includes: Processor and computer storage media; The processor is adapted to implement one or more instructions, the computer storage medium stores one or more instructions, and the one or more instructions are adapted to be loaded by the processor and executed as described in any one of claims 1-11.

14. A computer storage medium, characterized in that, The computer storage medium stores one or more instructions, which are adapted to be loaded by a processor and executed by the information processing method as described in any one of claims 1-11.

15. A computer program product, characterized in that, The computer program product includes one or more instructions; when one or more instructions in the computer program are executed by a processor, they implement the information processing method as described in any one of claims 1-11.

Images