Judgment document abstract method combining prompt learning and Qwen large language model

Abstract

The invention provides a judgment document method combining prompt learning and a Qwen large language model, which comprises the following steps of: extracting key feature words and corresponding contents of unique structure information in a judgment document data set by applying a text analysis method; by taking prompt learning as a carrier, constructing a prompt template containing training set unique structure information; combining the training set with the prompt template, and obtaining a reference large language model for the judgment document abstract task through fine tuning operation of the Qwen large language model; according to the method, the prompt template is combined with a newly accepted judgment document to be processed to construct a question-answer pair data format, and then the reference large language model is utilized to automatically generate a judgment document abstract, so that key information of the judgment document is not lost in an initial stage; the problem that some key information is lost due to text extraction in order to reduce the context length is avoided; and due to the assistance of the prompt template, the perception of key information of the judgment document can be further enhanced, and error problems are reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of large language model application, and in particular to a judicial document summary method combining prompt learning and a Qwen large language model. Background Art

[0002] With the deep integration of digital economy and smart justice, exploring the digital transformation of justice has become a hot topic of concern; however, against the background of increasing judicial needs, courts need to deal with a large number of judicial documents on a daily basis. Therefore, a concise and clear summary of judicial documents can enable judges to quickly grasp the case while maintaining efficiency and accuracy. It also makes it easier for judges, lawyers and other legal practitioners to quickly retrieve guiding cases from numerous cases.

[0003] The automated generation of judicial summaries has gradually become one of the core elements of judicial digital transformation, and plays a vital role in promoting the modernization of the trial system and promoting the development of judicial digitization and intelligence. In the judicial field, the automated generation of judicial summaries aims to summarize the contents of judicial documents in order to generate key information with fluent language and rigorous structure that can fully reflect the case trial process. However, due to the complexity of the information structure in judicial documents and the characteristics of ultra-long texts, the judicial summaries obtained by existing automated summary generation systems or methods are prone to key information loss, misleading subsequent judicial work and easily causing irreparable losses.

[0004] In order to overcome the shortcomings of existing automatic generation methods, some new technical means have emerged based on the characteristics of long texts of judicial documents with strong domain structures, such as the Chinese patent "CN117435725A", the invention name: A summary generation method and system for judicial judgment documents, which provides a summary generation method and system for judicial judgment documents; the method first divides the document into paragraphs based on the argument structure of the judicial judgment document, including the litigation request part, the defendant's defense part, the trial and investigation part and the judgment result part; then, the BERT model and the GCN model are combined to extract key sentences from each argument structure paragraph to form a key sentence set; next, the key sentence set is input into the generation model, and the generation model is used to rewrite the content to generate a sub-summary of each argument structure paragraph; finally, the sub-summaries corresponding to each argument structure are spliced ​​to form the final judicial judgment document summary.

[0005] However, the technical solution recorded in the patent may cause key information loss when extracting key sentences from the argument structure, thereby affecting the performance of the final summary. In addition, in the process of generating the final summary, the technical solution adopts a method of generating sub-summaries of each structural paragraph based on the key sentence set input generation model and then splicing them together. This method cannot fully perceive the logical relationship between the structures of each paragraph of the judgment document during the processing process, resulting in the lack of semantic relationship between the sub-paragraphs in the final generated summary.

[0006] At the same time, in the existing technology, the large language model's ability to perceive the internal structural information of the document and the complex logical relationships between the structures is relatively weak. This is because it lacks understanding of judicial logic and learning of the structure of judicial documents. As a result, when the model generates summaries or performs other related tasks, it may not be able to accurately reflect the core structural content and legal logic of judicial documents, thereby restricting its application effect in the judicial field.

[0007] Nowadays, large language models such as ChatGPT have begun to emerge. Based on rich knowledge base content, large language models have brought new opportunities for text summarization tasks with their excellent language understanding and text generation capabilities and achieved remarkable results. However, there are still gaps and considerable challenges in the research of large language models on text summarization tasks in the judicial field, that is, judicial document summary tasks. This is because the application of large language models to judicial document summary tasks faces the following two problems:

[0008] ① Lack of professional knowledge in the judicial field;

[0009] The current large language models focus on the research of general text analysis, and the exploration of vertical fields is slightly insufficient. They rarely involve tasks in the judicial field, such as summary generation tasks of judicial documents. Compared with general texts, judicial documents have unique legal terms and professional vocabulary. At present, emerging large language models are mostly trained based on general databases and lack the learning of professional knowledge in the judicial field. Therefore, when they are transferred to the task of summarizing judicial documents, they often expose the disadvantages of insufficient adaptability in professional fields, which is mainly reflected in the difficulty of the model to understand the task objectives, and the generated content is easy to deviate from reality or produce "hallucinations".

[0010] ②It is difficult to understand the structure and logic of judicial documents;

[0011] Judgment documents not only follow specific norms in the judicial field, but also have a strict logical structure and rigorous and formal language style. The production of summaries of judgment documents requires the model to have stronger text comprehension and logical reasoning capabilities to ensure that the key information in the text can be accurately captured. Summary of the invention

[0012] In order to solve the above technical problems, the present invention provides a judicial document summary method combining prompt learning and Qwen large language model, which overcomes the difficulty of processing long texts of judicial documents and optimizes the final generated judicial document summary to make it more in line with the language style of the judicial field, and can better reflect the logical structure and logical relationship between the various structures of the original text of the judicial document, thereby avoiding the loss of key information due to block summary to the greatest extent.

[0013] The judicial document method combining prompt learning and Qwen large language model includes:

[0014] Step 1: Use text parsing methods to extract key feature words and corresponding content of unique structural information in the judgment document dataset;

[0015] As an example, the unique structural information refers to: selecting the unique structural information in the judicial document dataset based on the authoritative dataset CALA2020.

[0016] As an example, the judicial document dataset is divided into: a training set, a validation set, and a test set.

[0017] As an example, the training set, validation set and test set all include: multiple judicial documents.

[0018] As an example, the text parsing method includes:

[0019] ① First, read the training set sentence by sentence to identify candidate feature keywords that can reflect its unique structural information;

[0020] As an example, the number of candidate feature keywords is: multiple groups.

[0021] ② On the basis of obtaining the candidate feature keywords, traverse any judicial document in the training set sentence by sentence, and once the current sentence in the judicial document contains the candidate feature keywords, determine the current sentence as the beginning of the current structure;

[0022] ③ Then continue to traverse in order until another candidate feature keyword is found, and classify all sentences between the two candidate feature keywords as the complete paragraph corresponding to the previous candidate feature keyword;

[0023] ④ And so on, until all judicial documents in the training set are traversed.

[0024] By traversing the training set, we can finally comprehensively and accurately obtain the key feature words and corresponding contents of the unique structural information of any judicial document in the training set.

[0025] Step 2: Using prompt learning as a carrier, construct a prompt template containing unique structural information of the training set;

[0026] First, provide a clear role positioning for the Qwen large language model, so that it knows that its main task is to process judicial document summaries and can follow instructions to process judicial documents in the template;

[0027] Secondly, based on the key feature words, construct instructions with precise language and clear purpose, further enhancing the understanding and application of instructions by the Qwen large language model;

[0028] Finally, according to the Prompt specification of the Qwen large language model, the candidate key feature words are combined with the Prompt specification of the Qwen large language model to construct a prompt template containing unique structural information of the training set, and the prompt template has distinct characteristics of judicial judgment documents;

[0029] As an example, the Qwen large language model can also adopt a general large language model similar to other models.

[0030] Step 3: Combine the training set with the prompt template, format it into a question-answer pair data format, and use it as the training set used for fine-tuning the Qwen large language model, that is, as the overall input of the Qwen large language model;

[0031] As an example, taking into full consideration the data characteristics displayed in the pre-training stage of the Qwen large language model, the question-answer pair data format can better bring into play the potential capabilities of the Qwen large language model; therefore, on top of the prompt template, the training set is combined with the prompt template, and this combination is constructed into multiple batches of single-round question-answer pair data formats as the overall input of the Qwen large language model.

[0032] Step 4: fine-tune the training set through the fine-tuning operation of the Qwen large language model, and then fine-tune the Qwen large language model through the SFT method, so that the Qwen large language model after the fine-tuning operation can be aligned with the judgment document summary task, and obtain a benchmark large language model for the judgment document summary task;

[0033] As an example, the SFT method is: Supervised Fine-Tuning.

[0034] As an example, using the question-answer data format and the LORA method to fine-tune the Qwen large language model can ensure that the benchmark large language model can greatly save computing resources while aligning the judicial document summary task.

[0035] Step 5: Use the prompt template in combination with the newly received judgment document to be processed to construct a question-answer pair data format, and then use the benchmark large language model to automatically generate a summary of the judgment document.

[0036] As an example, the step five is not limited to combining one newly accepted judgment document to be processed at a time, but can combine multiple newly accepted judgment documents to be processed at the same time, and automatically generate multiple comparative judgment document summaries.

[0037] Beneficial effects of the present invention:

[0038] Compared with the prior art, the technical solution proposed in the present invention enables the Qwen large language model to establish an overall perception of the full text of the judicial document, thereby ensuring that key information of the judicial document is not lost in the preliminary stage, and avoiding the problem of losing certain key information while extracting text in order to reduce the context length; and due to the assistance of the prompt template, the Qwen large language model can further enhance its perception of the key information of the judicial document, so that the generated judicial document summary is more in line with the language style of the judicial field, and the hallucination problem and error problem that occur when the Qwen large language model performs the generation task are alleviated. Judging from the experimental results and cases, the effect of the present invention is very good and basically meets the requirements of the judicial document summary task. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is a schematic diagram showing an example of Example 1 of the judicial document method combining prompt learning and the Qwen large language model of the present invention.

[0040] Figure 2 This is a reference schematic diagram for the process design of the judicial document method combining prompt learning and the Qwen large language model of the present invention.

[0041] Table 1 is a comparison table of experimental results generated by the judicial document method of the present invention combining prompt learning and Qwen large language model using the present invention and other multiple existing technologies.

[0042] Table 2 is a table of experimental results of the judicial document method of the present invention combining prompt learning and the Qwen large language model to achieve optimization of the performance of the Qwen large language model in a specific field.

[0043] Table 3 is a further case study analysis table of the judicial document method combining prompt learning and Qwen large language model of the present invention, which is an optimized experimental result of Table 2.

[0044] Table 4 is a table of manual evaluation results of the judicial document method of the present invention combining prompt learning and Qwen large language model. DETAILED DESCRIPTION

[0045] Below, reference Figure 1 to Figure 2As shown, the judicial document method combining prompt learning and Qwen large language model includes:

[0046] Step 1: Use text parsing methods to extract key feature words and corresponding content of unique structural information in the judgment document dataset;

[0047] As an example, the unique structural information refers to: selecting the unique structural information in the judicial document dataset based on the authoritative dataset CALA2020.

[0048] As an example, the judicial document dataset is divided into: a training set, a validation set, and a test set.

[0049] As an example, the training set, validation set and test set all include: multiple judicial documents.

[0050] As an example, the text parsing method includes:

[0051] ① First, read the training set sentence by sentence to identify candidate feature keywords that can reflect its unique structural information;

[0052] As an example, the number of candidate feature keywords is: multiple groups.

[0053] ② On the basis of obtaining the candidate feature keywords, traverse any judicial document in the training set sentence by sentence, and once the current sentence in the judicial document contains the candidate feature keywords, determine the current sentence as the beginning of the current structure;

[0054] ③ Then continue to traverse in order until another candidate feature keyword is found, and classify all sentences between the two candidate feature keywords as the complete paragraph corresponding to the previous candidate feature keyword;

[0055] ④ And so on, until all judicial documents in the training set are traversed.

[0056] By traversing the training set, we can finally comprehensively and accurately obtain the key feature words and corresponding contents of the unique structural information of any judicial document in the training set.

[0057] Step 2: Using prompt learning as a carrier, construct a prompt template containing unique structural information of the training set;

[0058] First, provide a clear role positioning for the Qwen large language model, so that it knows that its main task is to process judicial document summaries and can follow instructions to process judicial documents in the template;

[0059] Secondly, based on the key feature words, construct instructions with precise language and clear purpose, further enhancing the understanding and application of instructions by the Qwen large language model;

[0060] Finally, according to the Prompt specification of the Qwen large language model, the candidate key feature words are combined with the Prompt specification of the Qwen large language model to construct a prompt template containing unique structural information of the training set, and the prompt template has distinct characteristics of judicial judgment documents;

[0061] As an example, the Qwen large language model can also adopt a general large language model similar to other models.

[0062] Step 3: Combine the training set with the prompt template, format it into a question-answer pair data format, and use it as the training set used for fine-tuning the Qwen large language model, that is, as the overall input of the Qwen large language model;

[0063] As an example, taking into full consideration the data characteristics displayed in the pre-training stage of the Qwen large language model, the question-answer pair data format can better bring into play the potential capabilities of the Qwen large language model; therefore, on top of the prompt template, the training set is combined with the prompt template, and this combination is constructed into multiple batches of single-round question-answer pair data formats as the overall input of the Qwen large language model.

[0064] Step 4: fine-tune the training set through the fine-tuning operation of the Qwen large language model, and then fine-tune the Qwen large language model through the SFT method, so that the Qwen large language model after the fine-tuning operation can be aligned with the judgment document summary task, and obtain a benchmark large language model for the judgment document summary task;

[0065] As an example, the SFT method is: Supervised Fine-Tuning.

[0066] As an example, using the question-answer data format and the LORA method to fine-tune the Qwen large language model can ensure that the benchmark large language model can greatly save computing resources while aligning the judicial document summary task.

[0067] Step 5: Use the prompt template in combination with the newly received judgment document to be processed to construct a question-answer pair data format, and then use the benchmark large language model to automatically generate a summary of the judgment document.

[0068] As an example, the step five is not limited to combining one newly accepted judgment document to be processed at a time, but can combine multiple newly accepted judgment documents to be processed at the same time, and automatically generate multiple comparative judgment document summaries.

[0069] In order to better illustrate the specific implementation of the present invention, the processing method is described as follows through specific embodiments:

[0070] Embodiment 1:

[0071] Reference Figure 1 As shown:

[0072] First, through the judgment document dataset, we obtain the relevant structural information of the judgment document, namely S10;

[0073] Then, S11 is executed to construct a prompt template containing the structural information in S10 through prompt learning;

[0074] Next, S12 is executed to combine the judgment documents in the training set of the judgment document dataset with the prompt template and construct the data format of single-round multi-batch question-answer pairs;

[0075] Again, execute S13 and fine-tune the Qwen large language model through SFT (Supervised Fine-Tuning) to obtain a benchmark large language model for the judicial document summary task;

[0076] Finally, execute S14. After the method completes the first three steps, it can start to formally provide functions. When a request to generate a judgment document is received, the received judgment document is combined with the prompt template in S10 as the input of the benchmark large language model fine-tuned in S13, so that the output of the benchmark large language model is used as the summary text of the received judgment document.

[0077] It can be seen that Example 1 overcomes the difficulty of processing long texts of judicial documents, and optimizes the final generated summary of the judicial documents to make it more in line with the language style of the judicial field, and can better reflect the logical structure and logical relationship between the various structures of the original text of the judicial documents, and avoid the loss of key information due to block summaries to the greatest extent.

[0078] In order to better illustrate the design idea of ​​the present invention, the principle analysis is now introduced again as follows:

[0079] The present invention is different from the prior art in that:

[0080] ① Fine-tune the Qwen large language model:

[0081] A large language model for the judicial document summary task was trained: Since the pre-training stage of the Qwen large language model is based on a general database, its capabilities are shown to be universal and generalized, and it cannot be directly applied to downstream tasks in the judicial field. The present invention fine-tunes the Qwen large language model to align it with the judicial document summary task, so that the judicial document summary generated by the Qwen large language model is more aligned with the style of the judicial document field, without missing logical features, language style, and semantic relationships between key structures of the judicial document; with the help of the ultra-long context length of the Qwen large language model, the entire judicial document can be input into the model, avoiding the omission of key information, and it is generated as a whole by a large language model, which can improve the overall semantic fluency of the generated judicial document summary.

[0082] ② Tips for learning:

[0083] Constructing a prompt template containing structural information: Due to the extremely long text length of judicial documents, the present invention uses a prompt template containing structural information to enhance the Qwen large language model's perception of the structure of judicial documents, and obtains a benchmark large language model for the judicial document summary task. The generalization ability of the Qwen large language model is transformed into the ability of the benchmark large language model in the judicial field, focusing on the professional field of the judicial document summary task, so that the judicial document summary generated by the benchmark large language model can better present the key information of the judicial document.

[0084] As an example, the present invention proposes a method for generating judicial document summaries that combines prompt learning with the Qwen large language model, and introduces structural information into the judicial document summary task through the prompt learning method. First, in order to address the problem that the model lacks professional knowledge in the judicial field, a data-driven approach is adopted, and the training set part of the judicial document data set and its key information are used in the form of question-answer pairs to drive the Qwen large language model to align with the judicial document summary task, thereby enhancing the Qwen large language model's understanding of specific terms, concepts and legal logic in the judicial field. At the same time, in response to the problem of difficulty in understanding the structure and logic of judicial documents, this patent is inspired by prompt learning and designs a prompt template containing structural information and role instructions based on the SFT (Supervised Fine-Tuning) Qwen large language model to enhance the Qwen large language model's perception of the structure and logical relationship of judicial documents, and guide the Qwen large language model to more accurately capture the key information in the text, thereby better generating judicial document summaries that reflect the internal logic and structural information of judicial documents.

[0085] As an example, the present invention fine-tunes the Qwen large language model to align it with the judicial document summary task, so that the text generated by the Qwen large language model is more aligned with the style of the judicial document field, without missing logical features, language style and semantic relationships between key structures of the judicial document; with the help of the ultra-long context length of the Qwen large language model, the entire judicial document can be input into the Qwen large language model, avoiding the omission of key information, and being generated by a large language model as a whole, which can improve the overall semantic fluency of the generated judicial document summary. Based on the above situation, a prompt template containing structural information is used to enhance the Qwen large language model's perception of the structure of the judicial document, and a benchmark large language model for the judicial document summary task is obtained, which transforms the generalization ability of the Qwen large language model into the ability of the benchmark large language model in the judicial field, focusing on the professional field of the judicial document summary task, so that the judicial document summary generated by the benchmark large language model can better present the key information of the judicial document.

[0086] The scheme of the present invention is scientific, simple, easy to implement, safe and effective, suitable for promotion, and is mainly reflected in the following two aspects:

[0087] The present invention significantly improves the applicability of the (Qwen) large language model in the judicial field, making it more suitable for the needs of judicial document summary tasks;

[0088] As shown in Table 2, after using judicial document data to perform SFT fine-tuning on the Qwen large language model, the model can more accurately capture and understand the specific intent of the judicial document summary task, and then focus its generalization ability on the specific task of judicial document summary, thereby achieving the optimization of the Qwen large language model performance in a specific field.

[0089] In order to further enhance the sensitivity of the Qwen large language model to key information in judicial documents, this patent uses a prompt template containing structural information for enhancement. This method significantly improves the Qwen large language model's grasp of the language style in the judicial field when generating judicial document summaries, and effectively reduces the illusions and errors that often occur in the Qwen large language model in generation tasks;

[0090] In Table 2:

[0091] -SFT-Template: In this experiment, the Qwen-7B-Chat large language model did not use the SFT (Supervised Fine-Tuning) technology and the prompt template containing structural information. It only relied on the base Qwen-7B-Chat model for reasoning and then used automatic evaluation indicators to evaluate performance.

[0092] -SFT: In this experiment, the Qwen-7B-Chat large language model was not fine-tuned using the SFT technology. The performance was only evaluated using the automatic evaluation indicators after using the prompt template.

[0093] -Template: In this experiment, the Qwen-7B-Chat large language model is only fine-tuned through the SFT technology, without using prompt templates to enhance the large language model's perception of the structural information of judicial documents, and then the performance is evaluated using automatic evaluation indicators.

[0094] As shown in Table 1, the method proposed in this patent significantly surpasses other models in terms of automatic evaluation and summary generation, which preliminarily verifies the effectiveness of the method proposed in this patent;

[0095] In addition, as shown in Table 4, in the manual evaluation, experts prefer the summaries generated by this patent model;

[0096] Furthermore, as shown in Table 3, the left part of Table 3 is the summary of the judgment document obtained by the patented method, the middle part is the standard judgment document annotated by experts, and the right part is the judgment document generated by only fine-tuning the large language model by SFT without using prompt learning to enhance the Qwen large language model; it can be seen that the patented method not only fits the language logic style of the judicial field, but also strengthens the Qwen large language model's perception of the logical structure of the judgment document through prompt learning, thereby further reducing the hallucinations, errors and content loss problems that may be encountered when generating judgment document summaries.

[0097] Table 3 is a further case study of the optimization experimental results in Table 2, which is used to illustrate that the use of prompt learning can effectively improve the quality of generated judicial documents. The left and right sides are a comparison between the summaries obtained with and without prompt learning and the intermediate standard summary, which further illustrates the effectiveness of the method.

[0098] The above are only preferred embodiments of the present invention. It should be understood that the description of the above embodiments is only used to help understand the method and core ideas of the present invention, and is not used to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, etc. made within the ideas and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A judicial document method combining prompt learning and Qwen large language model, characterized in that: include: Step 1: Use text parsing methods to extract key feature words and corresponding content of unique structural information in the judgment document dataset; Step 2: Using prompt learning as a carrier, construct a prompt template containing unique structural information of the training set; First, provide a clear role positioning for the Qwen large language model, so that it knows that its main task is to process judicial document summaries and can follow instructions to process judicial documents in the template; Secondly, based on the key feature words, construct instructions with precise language and clear purpose, further enhancing the understanding and application of instructions by the Qwen large language model; Finally, according to the Prompt specification of the Qwen large language model, the candidate key feature words are combined with the Prompt specification of the Qwen large language model to construct a prompt template containing unique structural information of the training set, and the prompt template has distinct characteristics of judicial judgment documents; Step 3: Combine the training set with the prompt template, format it into a question-answer pair data format, and use it as the training set used for fine-tuning the Qwen large language model, that is, as the overall input of the Qwen large language model; Step 4: fine-tune the training set through the fine-tuning operation of the Qwen large language model, and then fine-tune the Qwen large language model through the SFT method, so that the Qwen large language model after the fine-tuning operation can be aligned with the judgment document summary task, and obtain a benchmark large language model for the judgment document summary task; Step 5: Use the prompt template in combination with the newly received judgment document to be processed to construct a question-answer pair data format, and then use the benchmark large language model to automatically generate a summary of the judgment document.

2. The judicial document method combining prompt learning and Qwen large language model according to claim 1 is characterized in that: The unique structural information refers to: based on the authoritative dataset CALA2020, the unique structural information in the judicial document dataset is selected.

3. The judicial document method combining prompt learning and Qwen large language model according to claim 1 is characterized in that: The judicial document dataset is divided into: training set, validation set, and test set.

4. The judicial document method combining prompt learning and Qwen large language model according to claim 1, characterized in that: The training set, validation set and test set all include: multiple judicial documents.

5. The judicial document method combining prompt learning and Qwen large language model according to claim 1, characterized in that: The text parsing method comprises: ① First, read the training set sentence by sentence to identify candidate feature keywords that can reflect its unique structural information; the number of the candidate feature keywords is: multiple groups; ② On the basis of obtaining the candidate feature keywords, traverse any judicial document in the training set sentence by sentence, and once the current sentence in the judicial document contains the candidate feature keywords, determine the current sentence as the beginning of the current structure; ③ Then continue to traverse in order until another candidate feature keyword is found, and classify all sentences between the two candidate feature keywords as the complete paragraph corresponding to the previous candidate feature keyword; ③ And so on, until all the judicial documents in the training set are traversed; By traversing the training set, we can finally comprehensively and accurately obtain the key feature words and corresponding contents of the unique structural information of any judicial document in the training set.

6. The judicial document method combining prompt learning and Qwen large language model according to claim 1, characterized in that: The Qwen large language model may also adopt other general large language models.

7. The judicial document method combining prompt learning and Qwen large language model according to claim 1 is characterized in that: The SFT method is: Supervised Fine-Tuning.

8. The judicial document method combining prompt learning and Qwen large language model according to claim 1, characterized in that: Fine-tuning the Qwen large language model using the question-answer data format and the LORA method can ensure that the benchmark large language model can greatly save computing resources while aligning the judicial document summary task.

9. The judicial document method combining prompt learning and Qwen large language model according to claim 1, characterized in that: The step five is not limited to combining one newly accepted judgment document to be processed at a time, but can combine multiple newly accepted judgment documents to be processed at the same time, and automatically generate multiple comparative judgment document summaries.

Images