Judicial field semantic abstracting method, system and equipment based on large model and storage medium

Abstract

The invention discloses a judicial field semantic abstracting method, system and device based on a large model and a storage medium. The method comprises the steps that voice recognition is conducted on a court trial video, voice is converted into a court trial text, the court trial text is labeled, and a labeled text set is obtained to serve as a training set; training the large model by using the labeled text set, and finely adjusting the large model after each round of training; performing voice recognition on the actual court trial video, converting the voice into an actual court trial text, and extracting a timestamp; utilizing the trained large model to identify the court hearing process of the actual court hearing text, and reasoning a court hearing process text; matching the reasoned court hearing process text to timestamps by using a matching algorithm, and generating video anchor points when the timestamps at the starting positions of all the processes in the court hearing process text correspond to the anchor points of the processes of the court hearing video; abstracting the reasoned court hearing process text; and combining the abstract text with the video anchor points to generate structured court hearing information. According to the invention, the court trial video processing efficiency is improved.

Description

Technical Field

[0001] The present invention relates to natural language processing technology, and in particular to a semantic summarization method, system, device and storage medium in the judicial field based on a large model. Background Art

[0002] As an important record of judicial procedures, the integrity and authenticity of the content of court trial videos are crucial for legal practitioners to understand cases and formulate strategies. With the development of technology, more and more court trial processes are recorded and stored, forming a huge video database. However, court trial videos are usually long and contain a large amount of information. If they rely on manual viewing one by one, it is not only inefficient, but also easy to miss key information.

[0003] At present, the processing of court trial videos mainly relies on manual operations, which is not only time-consuming and labor-intensive, but also difficult to be comprehensive and accurate due to the limited memory and attention of humans. For example, after the trial, the judge may need to review a specific trial link, such as "court investigation" or "court debate". If it relies on manual search, it will take a lot of time to fast forward and replay the video.

[0004] In addition, the analysis and summary of trial videos are of great value to judges writing judgments, lawyers preparing cases, and law school students studying. However, due to the lack of effective tools, these tasks often rely on text-based trial transcripts, which may not be able to fully record all the details of the trial process, such as the judge's tone of voice and the lawyer's debate skills.

[0005] In response to the above problems, although there are some open source and commercial speech recognition and text summarization models on the market, they often lack in-depth understanding of the specific field of trial. These models are usually trained based on general corpora and lack professional knowledge of legal terminology and trial process. As a result, when processing trial texts, they cannot accurately identify and segment the trial process, nor can they generate high-quality summaries that meet the needs of judges and lawyers.

[0006] Therefore, it is necessary to develop an intelligent processing system specifically for trial videos, which can automatically identify the various processes in the trial videos and generate structured summary information to improve the utilization efficiency of trial video materials. Summary of the invention

[0007] In view of the deficiencies in the prior art, the present invention provides a semantic summarization method, system, device and storage medium in the judicial field based on a large model, which can accurately identify each process in the trial video and perform high-quality speech summarization.

[0008] To achieve the above object, the present invention adopts the following technical solutions:

[0009] A semantic summarization method for judicial field based on a large model comprises the following steps:

[0010] Produce a data set, including: perform speech recognition on historical trial videos, convert speech into trial text, annotate the trial text to divide the trial process, and obtain annotated text set; divide the annotated text set into a training set and a validation set;

[0011] Use the training set to train the large model, and after each round of training, use the validation set to fine-tune the large model;

[0012] Perform speech recognition on the actual trial video, convert the speech into the actual trial text, and extract the timestamp;

[0013] Use the trained large model to identify the trial process of the actual trial text and infer the trial process text;

[0014] The inferred trial process text is matched to the timestamp using a matching algorithm. The timestamp of the start position of each process in the trial process text corresponds to the anchor point of the process in the trial video, generating a video anchor point.

[0015] Summarize the inferred trial process text and generate a summary text;

[0016] Combine summary text with video anchors to generate structured trial information.

[0017] To optimize the above technical solutions, the specific measures taken also include:

[0018] Furthermore, after each round of training, the large model is fine-tuned using the validation set as follows:

[0019] The validation set is used to evaluate the large model obtained in each round of training. The evaluation method is as follows: use the large model inference validation set obtained in the current round of training to calculate the inference accuracy, and select samples whose model inference results are different from the manual annotation results as negative samples, analyze the negative samples, confirm and correct these negative samples, or directly remove them from the training set and validation set.

[0020] Furthermore, the matching algorithm is a word-by-word matching algorithm. The specific process of using the word-by-word matching algorithm to match the inferred trial process text to the timestamp is as follows:

[0021] The actual trial text with timestamp is processed into a key-value map format, where the key is the timestamp and the value is the actual trial text corresponding to the timestamp. The first M words of the trial process text inferred by the large model are matched with the actual trial text. Starting from the first word of the trial process text, it is compared with each value in the actual trial text. Once it matches a word in a value in the actual trial text, it continues to match the second word. Otherwise, it continues to traverse the remaining words in the value and each remaining value; until the words in the inferred trial process text are completely matched with the words in the actual trial text, it is considered that the key corresponding to the value is the start timestamp of the trial process in the inferred trial process text.

[0022] Furthermore, the matching algorithm is a ROUGE fuzzy matching algorithm. The specific process of using the ROUGE fuzzy matching algorithm to match the inferred trial process text to the timestamp is as follows:

[0023] First, the actual trial text with timestamp is divided into sentences according to the three punctuation marks ".?!", and each sentence generates a corresponding key-value map form, the key is the timestamp, and the value is the actual trial text corresponding to the timestamp; the inferred trial process text is also divided into sentences according to the three punctuation marks ".?!", and the first three sentences of the inferred trial process text are taken, starting from the first sentence of the actual trial text, and compared with the first sentence of the inferred trial process text, and the similarity value is obtained through the calculation method of ROUGE-1. If the similarity value of the first sentence reaches the first threshold or above, the second sentence of the inferred trial process text is compared with the next sentence of the actual trial text. If the similarity of the second sentence reaches the second threshold or above, the third sentence is compared. If the similarity of the third sentence reaches the third threshold or above, it is considered that the match is successful and the timestamp is obtained; if the similarity of any of the sentences does not meet the standard, the match starts again from the next sentence of the actual trial text.

[0024] Furthermore, the similarity value obtained by the calculation method of ROUGE-1 is specifically:

[0025] All words were extracted from the actual trial text and the inferred trial process text;

[0026] Count the number of words that appear simultaneously in the actual trial text and the inferred trial process text;

[0027] The recall rate of ROUGE-1 is calculated using the following formula:

[0028]

[0029] In the formula, represents the recall rate of ROUGE-1, Indicates the number of words that appear simultaneously in the actual trial text and the inferred trial process text, represents the total number of words in the actual trial text;

[0030] The accuracy of ROUGE-1 is calculated using the following formula:

[0031]

[0032] In the formula, represents the accuracy of ROUGE-1, Indicates the number of words that appear simultaneously in the actual trial text and the inferred trial process text, The total number of words in the trial process text that represent reasoning;

[0033] The harmonic mean of precision and recall is used as the similarity value, and the calculation formula is as follows:

[0034]

[0035] In the formula, Represents the similarity value.

[0036] Furthermore, the summary of the inference trial process text is as follows:

[0037] Determine the length of the text of each trial process. If the length is the preset number of words or less, select the organization prompt words and rewrite the text of the trial process in written form;

[0038] If the length is longer than the preset number of words, the text of each trial process is extracted through prompt words, including: the plaintiff's statement of litigation request, the defendant's defense, the witness or expert testimony in court, the judge's summary of the focus of the dispute and undisputed facts, and the parties presenting evidence to the court;

[0039] The extracted elements are integrated into new prompt words, and the trial process text is summarized through a large model. Exclusive prompt words are designed for the summary of each trial process text, so that each trial process is adapted to the elements in each trial process text.

[0040] The present invention also proposes a semantic summary system in the judicial field based on a large model, comprising:

[0041] The data set preparation module is used to prepare the training set, including: performing speech recognition on the trial video, converting the speech into trial text, annotating the trial text to divide the trial process, and obtaining the annotated text set as the training set;

[0042] The training module is used to train the large model using the annotated text set and fine-tune the large model after each round of training;

[0043] The speech recognition module is used to perform speech recognition on the actual trial video, convert the speech into the actual trial text, and extract the timestamp;

[0044] The reasoning module is used to use the trained large model to identify the trial process of the actual trial text and infer the trial process text;

[0045] A matching module is used to match the inferred trial process text to the timestamp using a matching algorithm, where the timestamp of the start position of each process in the trial process text corresponds to the anchor point of the process in the trial video, thereby generating a video anchor point;

[0046] A summary module is used to summarize the inferred trial process text and generate a summary text;

[0047] The presentation module is used to combine summary text with video anchors to generate structured trial information.

[0048] The present invention also proposes an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the large model-based semantic summarization method for the judicial field as described above is implemented.

[0049] The present invention also provides a computer-readable storage medium storing a computer program, wherein the computer program enables a computer to execute the large model-based semantic summarization method for the judicial field as described above.

[0050] The beneficial effects of the present invention are:

[0051] (1) Automatic recognition of court trial process: Through the deep learning model, automatic recognition of court trial process is achieved without the need for human intervention, greatly improving the processing speed.

[0052] (2) Structured trial summary: The generated summary information is structured, which makes it easier for judges to quickly understand the key points of the trial and facilitates the preparation of documents and judgments.

[0053] (3) Accurate matching of video anchor points: Through innovative matching algorithms, the trial process is accurately linked to the video content, making it easier for judges to review the case.

[0054] (4) Ability to process long texts: The model solves the problem of long text processing and is able to process court trial texts of tens of thousands of words, thus improving the scope of application of the model.

[0055] (5) Integration of professional knowledge: By integrating legal professional knowledge, the model’s ability to understand and process trial texts is improved.

[0056] (6) Efficient use of training and validation sets: By establishing a labeling platform and optimizing model training, the utilization efficiency of training and validation sets is improved.

[0057] (7) Negative sample analysis and model adjustment: Through negative sample analysis and model adjustment, the accuracy and robustness of the model are improved.

[0058] (8) Verification of reasoning accuracy: By formulating a series of rules and scripts, the accuracy of model reasoning is verified and the credibility of the model is improved.

[0059] (9) Optimization of model inference results: Through continuous analysis and fine-tuning, the model’s compliance rate is improved and the quality of the processing results is ensured.

[0060] (10) User-friendly system design: The system provides flexible configurable parameters, and users can adjust parameters according to their needs to adapt to different application scenarios.

[0061] (11) The technical solution of the present invention not only solves the technical difficulties in court trial video processing, but also provides judges and legal workers with an efficient and accurate auxiliary tool. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 This is a flowchart of the semantic summarization method in the judicial field based on a large model proposed in the present invention.

[0063] Figure 2 This is a schematic diagram of the trial process. DETAILED DESCRIPTION

[0064] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0065] Example 1

[0066] The present invention proposes a semantic summarization method for judicial field based on a large model. The flowchart of the method is as follows: Figure 1 As shown, the following steps are included:

[0067] Produce a data set, including: perform speech recognition on historical trial videos, convert speech into trial text, annotate the trial text to divide the trial process, and obtain annotated text set; divide the annotated text set into a training set and a validation set;

[0068] The training set is used to train the large language model. This embodiment uses open source large models, such as Qwen1.5-14B, Qwen2-7B, Qwen2-14B of Alibaba Group, and GLM4 of Tsinghua University. After each round of training, the large model is fine-tuned using the validation set; according to the characteristics of the trial video, the training of the large model can obtain a vertical field model, which improves the accuracy of the trial process recognition. Specifically:

[0069] The large model obtained from each round of training is evaluated using the validation set. The evaluation method is: use the large model inference validation set obtained from the current round of training to calculate the inference accuracy, and screen out samples whose model inference results are different from those of the manual annotations as negative samples, analyze the negative samples, confirm and correct these negative samples, or directly remove them from the training set and validation set. This embodiment attempts fine-tuning training on large models such as Qwen1.5-14B, Qwen2-7B, Qwen2-14B, and GLM4, and repeatedly fine-tunes parameters such as batch size batch_size and learning rate learning_rate to ensure that the training loss converges as much as possible. Finally, Qwen2-7B, which has a better effect, is selected as the large model for trial process recognition in subsequent actual trial texts.

[0070] Perform speech recognition on the actual trial video, convert the speech into the actual trial text, and extract the timestamp;

[0071] Use the trained large model to identify the trial process of the actual trial text and infer the trial process text;

[0072] The matching algorithm is used to match the inferred trial process text to the timestamp. The timestamp of the starting position of each process in the trial process text corresponds to the anchor point of the process in the trial video, generating a video anchor point.

[0073] The matching algorithm is a word-by-word matching algorithm. The specific process of using the word-by-word matching algorithm to match the inferred trial process text to the timestamp is as follows:

[0074] The actual trial text with timestamp is processed into a key-value map format, where the key is the timestamp and the value is the actual trial text corresponding to the timestamp. The first M words of the trial process text inferred by the large model are matched with the actual trial text. Starting from the first word of the trial process text, it is compared with each value in the actual trial text. Once it matches a word in a value in the actual trial text, it continues to match the second word. Otherwise, it continues to traverse the remaining words in the value and each remaining value; until the words in the inferred trial process text are completely matched with the words in the actual trial text, it is considered that the key corresponding to the value is the start timestamp of the trial process in the inferred trial process text.

[0075] The matching algorithm may also be the ROUGE fuzzy matching algorithm. The specific process of using the ROUGE fuzzy matching algorithm to match the inferred trial process text to the timestamp is as follows:

[0076] First, the actual trial text with timestamp is divided into sentences according to the three punctuation marks ".?!", and each sentence generates a corresponding key-value map form, the key is the timestamp, and the value is the actual trial text corresponding to the timestamp; the inferred trial process text is also divided into sentences according to the three punctuation marks ".?!", and the first three sentences of the inferred trial process text are taken, starting from the first sentence of the actual trial text, and compared with the first sentence of the inferred trial process text, and the similarity value is obtained through the ROUGE-1 calculation method. Specifically:

[0077] All words were extracted from the actual trial text and the inferred trial process text;

[0078] Count the number of words that appear simultaneously in the actual trial text and the inferred trial process text;

[0079] The recall rate of ROUGE-1 is calculated using the following formula:

[0080]

[0081] In the formula, represents the recall rate of ROUGE-1, Indicates the number of words that appear simultaneously in the actual trial text and the inferred trial process text, represents the total number of words in the actual trial text;

[0082] The accuracy of ROUGE-1 is calculated using the following formula:

[0083]

[0084] In the formula, represents the accuracy of ROUGE-1, Indicates the number of words that appear simultaneously in the actual trial text and the inferred trial process text, The total number of words in the trial process text that represent reasoning;

[0085] The harmonic mean of precision and recall is used as the similarity value, and the calculation formula is as follows:

[0086]

[0087] In the formula, Represents the similarity value.

[0088] If the similarity value of the first sentence reaches the first threshold or above, the second sentence of the inferred trial process text is compared with the next sentence of the actual trial text. If the similarity of the second sentence reaches the second threshold or above, the third sentence is compared. If the similarity of the third sentence reaches the third threshold or above, the match is considered successful and the timestamp is obtained. If the similarity of any of the sentences does not meet the standard, the match starts again from the next sentence of the actual trial text.

[0089] Summarize the inferred trial process text and generate a summary text. Specifically, determine the text length of each trial process. If the length is a preset number of words or less, select an organization prompt word and rewrite the text of the trial process in written form. The preset number of words in this embodiment is 20 words.

[0090] If the length is above the preset word count, the prompt words are used to extract elements from each trial process text, including: the plaintiff's statement of litigation request, the defendant's defense, the witness or expert testimony, the judge's summary of dispute points and undisputed facts, and the parties' presentation of evidence to the court; this solves the problem of long text processing and can process trial texts of tens of thousands of words, thus increasing the scope of application of the model. Figure 2 shown.

[0091] The extracted elements are integrated into new prompt words, and the trial process text is summarized through a large model. Exclusive prompt words are designed for the summary of each trial process text, so that each trial process is adapted to the elements in each trial process text.

[0092] Combine summary text with video anchors to generate structured trial information. The generated summary information is structured, which makes it easier for judges to quickly understand the key points of the trial and facilitates the writing of documents and judgments. The precise matching of video anchors accurately links the trial process with the video content, making it easier for judges to review the case.

[0093] Example 2

[0094] The present invention proposes a semantic summarization system for judicial field based on a large model corresponding to the method of embodiment 1, comprising:

[0095] The data set preparation module is used to prepare the training set, including: performing speech recognition on the trial video, converting the speech into trial text, annotating the trial text to divide the trial process, and obtaining the annotated text set as the training set;

[0096] The training module is used to train the large model using the annotated text set and fine-tune the large model after each round of training;

[0097] The speech recognition module is used to perform speech recognition on the actual trial video, convert the speech into the actual trial text, and extract the timestamp;

[0098] The reasoning module is used to use the trained large model to identify the trial process of the actual trial text and infer the trial process text;

[0099] A matching module is used to match the inferred trial process text to the timestamp using a matching algorithm, where the timestamp of the start position of each process in the trial process text corresponds to the anchor point of the process in the trial video, thereby generating a video anchor point;

[0100] A summary module is used to summarize the inferred trial process text and generate a summary text;

[0101] The presentation module is used to combine summary text with video anchors to generate structured trial information.

[0102] The implementation methods of each module and module function in the system are completely consistent with the steps of the method in Example 1, so they will not be repeated here.

[0103] The present invention provides a user-friendly system design. The system provides flexible configurable parameters, and users can adjust the parameters according to requirements to adapt to different application scenarios.

[0104] Example 3

[0105] The present invention proposes an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the semantic summarization method for the judicial field based on a large model as described in Example 1 is implemented.

[0106] Example 4

[0107] The present invention provides a computer-readable storage medium storing a computer program, wherein the computer program enables a computer to execute the semantic summarization method for the judicial field based on a large model as described in Example 1.

[0108] In the embodiments disclosed in the present application, the computer storage medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or apparatus. The computer storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or equipment, or any suitable combination of the foregoing. More specific examples of computer storage media may include electrical connections based on one or more lines, portable computer disks, hard disks, random access memories (RAM), read-only memories (ROM), erasable programmable read-only memories (EPROM or flash memory), optical fibers, portable compact disk read-only memories (CDROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0109] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0110] The above are only preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should be regarded as the protection scope of the present invention.

Claims

1. A semantic summarization method for judicial domain based on a large model, characterized in that: The following steps are involved: Produce a data set, including: perform speech recognition on historical trial videos, convert speech into trial text, annotate the trial text to divide the trial process, and obtain annotated text set; divide the annotated text set into a training set and a validation set; Use the training set to train the large model, and after each round of training, use the validation set to fine-tune the large model; Perform speech recognition on the actual trial video, convert the speech into the actual trial text, and extract the timestamp; Use the trained large model to identify the trial process of the actual trial text and infer the trial process text; The inferred trial process text is matched to the timestamp using a matching algorithm. The timestamp of the start position of each process in the trial process text corresponds to the anchor point of the process in the trial video, generating a video anchor point. Summarize the inferred trial process text and generate a summary text; Combine summary text with video anchors to generate structured trial information.

2. The method for semantic summarization in the judicial field based on a large model as claimed in claim 1, characterized in that: After each round of training, the validation set is used to fine-tune the large model as follows: The validation set is used to evaluate the large model obtained in each round of training. The evaluation method is as follows: use the large model inference validation set obtained in the current round of training to calculate the inference accuracy, and select samples whose model inference results are different from the manual annotation results as negative samples, analyze the negative samples, confirm and correct these negative samples, or directly remove them from the training set and validation set.

3. The method for semantic summarization in the judicial field based on a large model as claimed in claim 1, characterized in that: The matching algorithm is a word-by-word matching algorithm. The specific process of using the word-by-word matching algorithm to match the inferred trial process text to the timestamp is as follows: The actual trial text with timestamp is processed into a key-value map format, where the key is the timestamp and the value is the actual trial text corresponding to the timestamp. The first M words of the trial process text inferred by the large model are matched with the actual trial text. Starting from the first word of the trial process text, it is compared with each value in the actual trial text. Once it matches a word in a value in the actual trial text, it continues to match the second word. Otherwise, it continues to traverse the remaining words in the value and each remaining value; until the words in the inferred trial process text are completely matched with the words in the actual trial text, it is considered that the key corresponding to the value is the start timestamp of the trial process in the inferred trial process text.

4. The method for semantic summarization in the judicial field based on a large model as claimed in claim 1, characterized in that: The matching algorithm is the ROUGE fuzzy matching algorithm. The specific process of using the ROUGE fuzzy matching algorithm to match the inferred trial process text to the timestamp is as follows: First, the actual trial text with timestamp is divided into sentences according to the three punctuation marks ".?!", and each sentence generates a corresponding key-value map form, where the key is the timestamp and the value is the actual trial text corresponding to the timestamp; the inferred trial process text is also divided into sentences according to the three punctuation marks ".?!", and the first three sentences of the inferred trial process text are taken, starting from the first sentence of the actual trial text, and compared with the first sentence of the inferred trial process text, and the similarity value is obtained through the calculation method of ROUGE-1. If the similarity value of the first sentence reaches the first threshold or above, the second sentence of the inferred trial process text is compared with the next sentence of the actual trial text. If the similarity of the second sentence reaches the second threshold or above, the third sentence is compared. If the similarity of the third sentence reaches the third threshold or above, it is considered that the match is successful and the timestamp is obtained; if the similarity of any of the sentences does not meet the standard, the match starts again from the next sentence of the actual trial text.

5. The method for semantic summarization in the judicial field based on a large model as claimed in claim 4, characterized in that: The similarity value obtained by the ROUGE-1 calculation method is specifically: All words were extracted from the actual trial text and the inferred trial process text; Count the number of words that appear simultaneously in the actual trial text and the inferred trial process text; The recall rate of ROUGE-1 is calculated using the following formula: ； In the formula, represents the recall rate of ROUGE-1, Indicates the number of words that appear simultaneously in the actual trial text and the inferred trial process text, represents the total number of words in the actual trial text; The accuracy of ROUGE-1 is calculated using the following formula: ； In the formula, represents the accuracy of ROUGE-1, Indicates the number of words that appear simultaneously in the actual trial text and the inferred trial process text, The total number of words in the trial process text that represent reasoning; The harmonic mean of precision and recall is used as the similarity value, and the calculation formula is as follows: ； In the formula, Represents the similarity value.

6. The method for semantic summarization in the judicial field based on a large model as claimed in claim 1, characterized in that: The summary of the inference trial process text is as follows: Determine the length of the text of each trial process. If the length is the preset number of words or less, select the organization prompt words and rewrite the text of the trial process in written form; If the length is longer than the preset number of words, the text of each trial process is extracted through prompt words, including: the plaintiff's statement of litigation request, the defendant's defense, the witness or expert testimony in court, the judge's summary of the focus of the dispute and undisputed facts, and the parties presenting evidence to the court; The extracted elements are integrated into new prompt words, and the trial process text is summarized through a large model. Exclusive prompt words are designed for the summary of each trial process text, so that each trial process is adapted to the elements in each trial process text.

7. A semantic summarization system in the judicial field based on a large model, characterized in that: include: The data set preparation module is used to prepare the training set, including: performing speech recognition on the trial video, converting the speech into trial text, annotating the trial text to divide the trial process, and obtaining the annotated text set as the training set; The training module is used to train the large model using the annotated text set and fine-tune the large model after each round of training; The speech recognition module is used to perform speech recognition on the actual trial video, convert the speech into the actual trial text, and extract the timestamp; The reasoning module is used to use the trained large model to identify the trial process of the actual trial text and infer the trial process text; A matching module is used to match the inferred trial process text to the timestamp using a matching algorithm, where the timestamp of the start position of each process in the trial process text corresponds to the anchor point of the process in the trial video, thereby generating a video anchor point; A summary module is used to summarize the inferred trial process text and generate a summary text; The presentation module is used to combine summary text with video anchors to generate structured trial information.

8. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method for semantic summarization in the judicial field based on a large model as described in any one of claims 1 to 6 is implemented.

9. A computer-readable storage medium storing a computer program, characterized in that: The computer program enables the computer to execute the semantic summarization method for the judicial field based on a large model as described in any one of claims 1 to 6.

Images