Court trial record regularizing method, system, electronic device and product

Abstract

The application provides a court trial transcript regularization method and system, an electronic device and a product. The method performs noise reduction processing on a court trial transcription text to obtain a first processed text, and performs semantic purification on the court trial transcription text to obtain a second processed text. According to a comparison result of the first processed text and the second processed text, an initial regularization text corresponding to the court trial transcription text is determined. The initial regularization text is normalized to obtain a court trial standard text. Based on the court trial standard text and case metadata corresponding to the court trial transcription text, a court trial transcript corresponding to the court trial transcription text is determined. The scheme automatically regularizes the court trial transcription text, improves the production efficiency of the court trial transcript, and decouples the noise reduction processing and the semantic purification of the court trial transcription text into two independent tracks for parallel processing. The two track outputs are compared, and the initial regularization text is determined based on the comparison result. The scheme can avoid the situation that the text processing of a single track loses text content due to misjudgment, and improves the accuracy of the court trial transcript.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method, system, electronic device, and product for organizing court transcripts. Background Technology

[0002] Court transcripts, also known as trial records or trial records, are crucial documents recording court proceedings during judicial trials. They play a vital role in ensuring the fairness, accuracy, and traceability of trials. Traditionally, court transcripts were recorded manually by court clerks. However, with the continuous increase in the number of cases and judicial workload, relying solely on manual transcription by court clerks is no longer sufficient to meet the demands of efficiency and scale.

[0003] Currently, most courts have deployed real-time transcription systems in their trial courtrooms. These systems automatically generate transcribed texts with role tags during trials using speech recognition technology, enabling the input of trial transcripts. However, these transcribed texts are merely word-for-word records of spoken language in court, containing numerous interjections, repetitions, filler words, colloquial terms, and non-standard expressions. They fall far short of formal trial transcripts that conform to judicial document standards and can be directly archived and used. Court clerks primarily rely on manual transcription and proofreading of the transcribed texts after the trial before they can be used as trial transcripts, resulting in lengthy processing times and impacting the efficiency of trial transcript production. Summary of the Invention

[0004] Based on the above needs, this application proposes a method, system, electronic device, and product for standardizing court transcripts. This method can automatically standardize transcribed court transcripts, improve the efficiency of transcript production, and decouple the noise reduction preprocessing and semantic purification of the transcribed court transcripts into two independent parallel tracks. The outputs of the two tracks are then compared, and the initial standardized text is determined based on the comparison results. This avoids the loss of text content due to misjudgment in single-track text processing, thereby improving the accuracy of court transcripts.

[0005] To achieve the above objectives, this application proposes the following technical solution: According to a first aspect of the embodiments of this application, a method for organizing court hearing transcripts is provided, including: The court transcript was denoised to obtain the first processed text, and the court transcript was semantically purified to obtain the second processed text. Based on the comparison results between the first processed text and the second processed text, the initial normalized text corresponding to the court transcript is determined; The initial standardized text is then processed to obtain the standardized court hearing text; Based on the case metadata corresponding to the court hearing transcript and the court hearing transcription text, the court hearing record corresponding to the court hearing transcription text is determined.

[0006] Optionally, the noise reduction process performed on the transcribed court transcript to obtain the first processed text includes: The court transcript is subjected to spoken noise reduction, and / or, transcription error correction is performed on the court transcript, and / or, text compression is performed on the court transcript to obtain the first processed text.

[0007] Optionally, the semantic purification of the transcribed court transcript to obtain the second processed text includes: According to the pre-set semantic purification content, the court transcript is semantically purified to obtain a second processed text. The semantically purified content includes at least one of the following: core claims, evidence, cross-examination opinions, and key points of defense.

[0008] Optionally, based on the comparison result between the first processed text and the second processed text, the initial normalized text corresponding to the court transcript is determined, including: The first processed text and the second processed text are aligned to obtain at least one group of speech segments; the group of speech segments includes a first speech segment in the first processed text and a second speech segment in the second processed text, and the first speech segment and the second speech segment are mutually aligned speech segments; The first and second speech segments in the speech segment group are compared to obtain the comparison result corresponding to the speech segment group; Based on the comparison results corresponding to the speech segment group, the target speech segment corresponding to the speech segment group is determined; The target speech segments corresponding to the speech segment groups are combined in the order of the speech segment groups to obtain the initial regularized text corresponding to the court transcript.

[0009] Optionally, the first and second speech segments in the speech segment group are compared to obtain the comparison result corresponding to the speech segment group, including: The degree of divergence between the first and second speech segments is calculated based on the edit distance and / or semantic similarity between the first and second speech segments in the speech segment group; If the degree of divergence is greater than a preset degree of divergence threshold, then the comparison result corresponding to the speech segment group indicates that there is a significant divergence. If the degree of divergence is not greater than a preset degree of divergence threshold, then the comparison result corresponding to the speech segment group indicates that there is no significant divergence.

[0010] Optionally, based on the comparison results corresponding to the speech segment group, the target speech segment corresponding to the speech segment group is determined, including: If the comparison results corresponding to the speech segment group indicate a significant discrepancy, then the speech segment in the court transcript that corresponds to the first speech segment is taken as the target speech segment corresponding to the speech segment group, and the target speech segment is marked as content to be reviewed; If the comparison results corresponding to the speech segment group indicate that there is no significant disagreement, then the first speech segment is taken as the target speech segment corresponding to the speech segment group.

[0011] Optionally, the initial standardized text is subjected to normalization processing to obtain the standardized court hearing text, including: Based on a pre-constructed knowledge graph of courtroom roles, the initial standardized text is subjected to legal writing standardization processing to obtain a written courtroom text, which is then used as the standard courtroom text. The legal writing standardization processing includes: title replacement, application of courtroom stage templates, and standardization of legal terminology. The nodes in the knowledge graph of courtroom roles include: roles, legal titles, courtroom stage templates, and a legal terminology database.

[0012] Optionally, the initial standardized text is further processed to obtain the standardized court hearing text, including: The court hearing text is reconstructed according to a pre-set fact chain to obtain a standard court hearing text; wherein the fact chain includes at least one of the following: an event chronology chain, a causal relationship chain, and a fact logic chain.

[0013] Optionally, the written court hearing text is reconstructed according to a pre-set fact chain to obtain a standard court hearing text, including: According to preset extraction elements, the written text of the court hearing is subjected to element extraction to obtain the text elements in the written text of the court hearing; the preset extraction elements include at least one of the following: time, place, people, events, claims, evidence, cross-examination, debate, and conclusion; Based on the text elements and the pre-set fact chain, the speech segments in the written court text are reconstructed to obtain at least one reconstructed paragraph. All reconstructed paragraphs are then combined to obtain the standard court text.

[0014] Optionally, based on the case metadata corresponding to the court hearing transcript and the court hearing transcript, the court hearing record corresponding to the court hearing transcript is determined, including: According to the pre-set court transcript template, the case metadata corresponding to the transcribed court transcript and the standard court transcript are combined to obtain the initial transcript; Based on a pre-set library of standardized court transcript templates, the initial transcript is verified. If the verification passes, the initial transcript is used as the court transcript corresponding to the transcribed court transcript.

[0015] Optionally, determining the court transcript corresponding to the court transcript based on the case metadata corresponding to the court transcript and the court transcript text further includes: If the verification fails, the speech segments that fail the verification in the initial transcript are marked as content to be reviewed, and the error type of the content to be reviewed is determined. Based on the error type, the content to be reviewed is reorganized to obtain the reorganized content of the content to be reviewed; Based on the reorganized content and the initial transcript, the court transcript corresponding to the transcribed court transcript is determined.

[0016] According to a second aspect of the embodiments of this application, a court hearing transcript generation system is provided, including: a court hearing transcription system and a court hearing transcript formatting system; The court hearing transcription system is used to transcribe the speech during the court hearing to obtain the transcribed text, and then transmit the transcribed text to the court hearing transcript system. The court transcript formatting system is used to execute the aforementioned court transcript formatting method.

[0017] According to a third aspect of the embodiments of this application, an electronic device is provided, including: a memory and a processor; The memory is connected to the processor and is used to store programs; The processor is used to implement the above-mentioned method for organizing court transcripts by running the program in the memory.

[0018] According to a fourth aspect of the embodiments of this application, a computer program product is provided, including computer program instructions, which, when executed by a processor, cause the processor to implement the above-described court transcript organization method.

[0019] The court transcript standardization method proposed in this application involves denoising the transcribed court transcript to obtain a first processed text, semantically refining the transcribed text to obtain a second processed text, comparing the first and second processed texts to determine the initial standardized text corresponding to the transcribed text, standardizing the initial standardized text to obtain a standardized court transcript, and determining the corresponding court transcript based on the case metadata corresponding to the standardized court transcript and the transcribed text. Using the technical solution of this application, court transcripts can be automatically standardized, improving the efficiency of court transcript production. Furthermore, the denoising and semantic refining processes are decoupled into two independent parallel tracks, and the outputs of the two tracks are compared to determine the initial standardized text. This avoids the loss of text content due to misjudgment in single-track text processing, thus improving the accuracy of court transcripts. Attached Figure Description

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

[0021] Figure 1 This is a flowchart illustrating a method for organizing court hearing transcripts, provided as an embodiment of this application.

[0022] Figure 2 This is a schematic diagram of a standardized processing flow provided in an embodiment of this application.

[0023] Figure 3 This is a schematic diagram of another standardized processing flow provided in an embodiment of this application.

[0024] Figure 4 This is a schematic diagram illustrating a process for determining court hearing transcripts, provided as an embodiment of this application.

[0025] Figure 5 This is a schematic diagram illustrating another process for determining court transcripts, provided as an embodiment of this application.

[0026] Figure 6 This is a schematic diagram of a court transcript organizing device provided in an embodiment of this application.

[0027] Figure 7 This is a schematic diagram of the structure of a court hearing transcript generation system provided in an embodiment of this application.

[0028] Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0029] The technical solution of this application is applicable to the application scenario of generating court transcripts. By adopting the technical solution of this application, the transcribed text of court proceedings can be automatically standardized, improving the efficiency of court transcript production. Furthermore, the noise reduction preprocessing and semantic purification of the transcribed text are decoupled into two independent tracks for parallel processing. The outputs of the two tracks are then compared, and the initial standardized text is determined based on the comparison results. This avoids the loss of text content due to misjudgment in single-track text processing, thus improving the accuracy of court transcripts.

[0030] Court transcripts, also known as trial records or trial records, are judicial documents created by court clerks during litigation, truthfully recording the court proceedings. They not only objectively recreate the entire trial process but also serve as crucial evidence for drafting first-instance judgments, remanding cases for retrial, or reversing judgments in second-instance proceedings. They possess irreplaceable evidentiary value in appeals and retrials. Court transcripts are one of the fundamental sources for establishing facts in judgments. Their accuracy, completeness, and standardization directly affect the legality of the judgment. Any omission or distortion of key information can lead to a remand for retrial, seriously impacting judicial credibility.

[0031] In the traditional model, court transcripts rely entirely on court clerks to manually listen and type them. This method was sustainable when the number of cases was low, but with the continuous increase in the number of cases in recent years, relying solely on court clerks to manually transcribe has become insufficient to meet the demands of efficiency and scale.

[0032] To alleviate the aforementioned pressure, most intermediate and grassroots courts have now deployed real-time transcription systems in their trial courtrooms. These systems, based on Automatic Speech Recognition (ASR) and Speaker Diarization (SDI) technologies, can convert speech into text streams with role tags (such as "Presiding Judge," "Plaintiff," "Defendant," and "Witness") in real time during trials, essentially achieving full recording of trial content.

[0033] However, the transcripts output by the court hearing transcription system are essentially rough word-for-word records of spoken language in court. They contain a large number of interjections, repetitive sentences, filler words, colloquial terms, and non-standard expressions. They are still far from meeting the standards of judicial documents and can be directly archived and used as formal court hearing transcripts. Therefore, the transcripts still need to be manually transcribed and proofread by the court clerk after the hearing before they can be used as court hearing transcripts. This results in a long time to prepare the court hearing transcripts and affects the efficiency of court hearing transcript production.

[0034] Based on this, this application proposes a method for standardizing court transcripts. This technical solution can automatically standardize the transcribed text of court hearings, improve the efficiency of court transcript production, and decouple the noise reduction preprocessing and semantic purification of the transcribed text of court hearings into two independent tracks for parallel processing. The outputs of the two tracks are then compared, and the initial standardized text is determined based on the comparison results. This can avoid the loss of text content due to misjudgment in single-track text processing, thereby improving the accuracy of court transcripts.

[0035] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0036] Exemplary methods See Figure 1 As shown in the embodiment of this application, a method for organizing court hearing transcripts is proposed. The method includes: S101. Noise reduction is performed on the transcribed court transcript to obtain the first processed text. Semantic purification is then performed on the transcribed court transcript to obtain the second processed text.

[0037] A court hearing transcription system is used to perform speech recognition and voiceprint separation on the voices of various roles captured during the court hearing. The speech is transcribed into text, and corresponding role tags are generated based on the different voiceprints, resulting in a transcribed court hearing text with role tags. These role tags can include: judge, plaintiff, defendant, agent, witness, etc. Additionally, the transcribed court hearing text can include timestamps and clause boundaries.

[0038] This embodiment requires acquiring the court transcript output by the court transcript transcription system and performing noise reduction and semantic purification on the transcript. Noise reduction yields the first processed text, while semantic purification yields the second processed text. Noise reduction removes redundant noise and corrects transcription errors, such as removing modal particles, repetitive sentences, and filler words, and correcting errors in legal terminology, names, amounts, and dates. Semantic purification extracts core content from the transcript, achieving semantic-level compression.

[0039] In this embodiment, the court transcript is processed in parallel on two tracks: noise reduction and semantic purification. The first track performs noise reduction on the court transcript, and the second track performs semantic purification on the court transcript. The first track outputs the first processed text, and the second track outputs the second processed text.

[0040] In one specific implementation, the transcribed court transcript is subjected to noise reduction processing to obtain a first processed text, which specifically includes the following steps: The court transcript is subjected to spoken noise reduction, and / or, transcription error correction, and / or, text compression to obtain the first processed text.

[0041] In this embodiment, the noise reduction process includes at least one of spoken language noise reduction, transcription recognition error correction, and text expression compression.

[0042] When denoising includes spoken language denoising, the court transcript is denoised to remove redundant content, such as interjections, repetitive sentences, and filler words. In this embodiment, spoken language denoising can employ rule-based filtering, constructing a filter vocabulary and using regular expression matching to filter out content in the court transcript that matches words in the filter vocabulary; alternatively, it can employ classification labeling filtering, training a binary classifier using sample text pre-labeled with "validity" tags, and using this binary classifier to determine whether the content in the court transcript is valid, filtering out invalid content; or it can employ a large language model, setting a prompt that requires "deleting interjections, repetitive sentences, and filler words while retaining complete semantics," inputting the court transcript into this large language model, and performing spoken language denoising based on the prompt.

[0043] When noise reduction processing includes transcription error correction, the court transcript is transcribed to correct transcription errors, such as errors in legal terms, names, amounts, and dates. In this embodiment, transcription error correction can be achieved through rule-based correction. A confusion dictionary is constructed, including a mapping between incorrect and correct content (e.g., not following the right of defense → right of insecurity defense, etc.). Regular expression matching is used to replace content in the transcribed court text that matches the incorrect content in the confusion dictionary with the corresponding correct content in the confusion dictionary. Alternatively, a training error correction model can be used. This model is trained on pre-labeled error sample texts with correct content tags, with the training objective being that the predicted correct content output by the error correction model matches the correct content tags carried by the error sample texts. The trained error correction model is then used to correct the transcribed court text, outputting the corrected text. The error correction model can employ neural machine translation (NMT) models, end-to-end deep learning models (Seq2Seq models), etc. Finally, a large language model can be used, with a prompt indicating "correct the recognition errors while maintaining the original meaning." The transcribed court text is input into this large language model, and transcription error correction is performed based on the prompt.

[0044] In the noise reduction process, including text compression, the court transcript is compressed to extract all key points, thus compressing the text content. In this embodiment, text compression can be achieved by training a text compression model. This involves training a pre-built text compression model using sample text pre-labeled with compressed text tags, with the training objective being that the predicted compressed text output by the model matches the compressed text tags. The text compression model can be a TextRank model, a BERT model, or similar. Alternatively, a large language model can be used, with a prompt requiring "retaining all key points and removing redundancy." The court transcript is then input into this large language model, and text compression is performed based on the prompt.

[0045] In this embodiment, when the noise reduction process includes multiple components, the court transcript can be processed first, followed by a second processing step, and so on, until all noise reduction processes are completed. The final generated text is then used as the first processed text. For example, when the noise reduction process includes spoken language noise reduction, transcription error correction, and text compression, the court transcript can be first denoised to obtain a denoised text. Then, transcription error correction can be performed on the denoised text to obtain a corrected text. Finally, text compression can be performed on the corrected text to obtain a compressed text, which is then used as the first processed text. Alternatively, this embodiment can pre-train a noise reduction model that includes a spoken language noise reduction module, a transcription error correction module, and a text compression module. Each module can be trained according to the training methods described above. The trained noise reduction model can then be used to perform spoken language noise reduction, transcription error correction, and text compression on the court transcript to obtain the first processed text. This embodiment can also utilize a large language model, setting prompts that include spoken language noise reduction, transcription error correction, and text expression compression. The court transcript is input into the large language model, and noise reduction is performed on the court transcript based on the prompts.

[0046] In one specific implementation, semantic purification is performed on the transcribed court transcript to obtain a second processed text, specifically including the following steps: Based on the pre-set semantic purification content, the court transcript is semantically purified to obtain the second processed text.

[0047] This embodiment pre-sets semantic purification content. The court transcript is then semantically purified according to this content to obtain a second processed text. The semantically purified content includes at least one of the following: core claims, evidence, cross-examination opinions, and key points of the defense. In this embodiment, semantic purification can be achieved by training a semantic purification model. The model is trained using sample text labeled with purified text tags containing the semantically purified content, with the goal of ensuring that the predicted purified text output by the semantic purification model matches the purified text tags. The trained semantic purification model is used to perform semantic purification on the court transcript, resulting in a second-processed text. The semantic purification model can be an NLP model, or a large language model. A semantic purification prompt can be set, and the court transcript is input into this large language model. Semantic purification is then performed based on the prompt. For example, the prompt could be set to: "You are a court transcript compilation assistant. Please extract information from the following text according to the specified fields and output it in JSON format. Fields not mentioned should be filled with null. Fields: {Core Claims, Evidence, Cross-examination Opinions, Key Points of Reply}".

[0048] In this embodiment, the semantic purification is more effective than the text compression in the noise reduction process. In other words, the text content obtained after semantic purification is more concise than the text content obtained after text compression.

[0049] In addition, this embodiment can also set up a large language model with two-track parallel processing capability. The court transcript is input into the large language model. The first track module of the large language model performs noise reduction processing on the court transcript and outputs the first processed text. The second track module of the large language model performs semantic purification on the court transcript and outputs the second processed text. The first track module and the second track module of the large language model are respectively set with corresponding prompts. The setting of the prompts is the same as that described above, and will not be described in detail in this embodiment.

[0050] S102. Based on the comparison results of the first processed text and the second processed text, determine the initial standardized text corresponding to the court transcript.

[0051] In this embodiment, after obtaining the first processed text and the second processed text, it is necessary to compare the first processed text and the second processed text. This comparison requires alignment; that is, the first processed text corresponding to the text within a preset window area in the court transcript needs to be compared with the second processed text corresponding to the text within the same preset window area to determine the comparison result. Based on the comparison result, the initial standardized text corresponding to the court transcript is determined. Specifically, based on the comparison result for each preset window area, the initial standardized text corresponding to each preset window area is determined. Then, according to the order of the preset window areas, the initial standardized text corresponding to each preset window area is sorted and merged to obtain the initial standardized text corresponding to the court transcript.

[0052] In this embodiment, the comparison results include two types: significant discrepancy and no significant discrepancy. If the comparison result corresponding to the preset window area indicates significant discrepancy, the text within the preset window area in the court transcript is used as the initial normalized text corresponding to that preset window area. If the comparison result corresponding to the preset window area indicates no significant discrepancy, the first processed text corresponding to the text within that preset window area in the court transcript is directly used as the initial normalized text corresponding to that preset window area. When the comparison result indicates significant discrepancy, it indicates a large difference between the denoised text and the semantically purified text, possibly due to errors in the denoising or semantic purification processes. To avoid losing text content, the first processed text after denoising or the second processed text after semantic purification is no longer used as the initial normalized text; instead, the original content in the court transcript is directly used as the initial normalized text. When the comparison results show no significant discrepancy, it indicates that the difference between the denoised text and the semantically purified text is not significant, and there are no errors in the denoising and semantic purification processes. In this case, the first processed text after denoising is directly used as the initial normalized text. The reason for using the first processed text as the initial normalized text instead of the second processed text is that the second processed text is semantically purified, while the first processed text is only denoised. Compared with the second processed text, the content in the first processed text is more complete, so as to make the generation of the court transcript more accurate.

[0053] In one specific implementation, the initial standardized text corresponding to the court transcript is determined based on the comparison result of the first processed text and the second processed text, specifically including the following steps: First, the first and second processed texts are aligned in terms of speech segments to obtain at least one set of speech segment groups.

[0054] In this embodiment, a preset window area can be set as a speech segment, and the first processed text and the second processed text can be aligned to obtain at least one group of speech segments. That is, the first processed text and the second processed text corresponding to each speech segment in the court transcript are aligned. In this embodiment, each speech segment in the court transcript can be assigned a speech segment identifier. When the court transcript is denoised, the first speech segment after denoising each speech segment carries the corresponding speech segment identifier. Then, the first processed text contains at least one first speech segment with a speech segment identifier. When the court transcript is semantically purified, the second speech segment after semantic purification each speech segment carries the corresponding speech segment identifier. Then, the second processed text contains at least one second speech segment with a speech segment identifier.

[0055] Speech segments with the same speech segment identifier in the first processed text and the second processed text are divided into a speech segment group. That is, the speech segment group includes the first speech segment in the first processed text and the second speech segment in the second processed text, and the first speech segment and the second speech segment are mutually aligned speech segments, that is, speech segments with the same speech segment identifier.

[0056] Second, the first and second speaking segments in the speaking segment group are compared to obtain the corresponding comparison results for the speaking segment group.

[0057] This embodiment compares the first and second speech segments in a speech segment group to obtain the comparison result for the speech segment group. This embodiment can determine the comparison result for the speech segment group by calculating the edit distance between the first and second speech segments. When the edit distance is greater than a preset edit distance threshold, the comparison result indicates a significant discrepancy; when the edit distance is not greater than the preset edit distance threshold, the comparison result indicates no significant discrepancy. This embodiment can also determine the comparison result for the speech segment group by calculating the semantic similarity between the first and second speech segments. When the semantic similarity is greater than a preset similarity threshold, the comparison result indicates no significant discrepancy; when the semantic similarity is not greater than the preset similarity threshold, the comparison result indicates a significant discrepancy. The semantic similarity calculation can utilize a semantic model to determine the first semantic feature of the first speech segment and the second semantic feature of the second speech segment, calculate the cosine similarity between the first and second semantic features, and use this cosine similarity as the semantic similarity.

[0058] Specifically, the following steps can be used to compare the first and second speech segments within a speech segment group to obtain the corresponding comparison results for the speech segment group: First, the degree of divergence between the first and second speech segments is calculated based on the edit distance and / or semantic similarity between the first and second speech segments in the speech segment group.

[0059] This embodiment can calculate the edit distance between the first and second speech segments in a speech segment group, normalize the edit distance to obtain the normalized edit distance, and use the difference between 1 and the normalized edit distance (i.e., 1 - normalized edit distance) as the degree of divergence between the first and second speech segments.

[0060] This embodiment can also calculate the semantic similarity between the first and second speech segments in the speech segment group, and use the difference between 1 and the semantic similarity (i.e., 1 - semantic similarity) as the degree of divergence between the first and second speech segments. The method for calculating semantic similarity has been described above.

[0061] In this embodiment, a first weight and a second weight can be preset, wherein the first weight is the weight of the normalized edit distance, the second weight is the weight of the semantic similarity, the sum of the first weight and the second weight is 1, and 1 - (first weight × normalized edit distance + second weight × semantic similarity) is used as the degree of divergence between the first speech segment and the second speech segment.

[0062] Secondly, if the degree of disagreement is greater than the preset degree of disagreement threshold, it is determined that the comparison result corresponding to the speech segment group indicates significant disagreement; if the degree of disagreement is not greater than the preset degree of disagreement threshold, it is determined that the comparison result corresponding to the speech segment group indicates no significant disagreement.

[0063] This embodiment pre-sets a preset divergence threshold. If the divergence between the first and second speech segments is greater than the preset divergence threshold, the comparison result corresponding to the speech segment group is determined to indicate significant divergence. If the divergence between the first and second speech segments is not greater than the preset divergence threshold, the comparison result corresponding to the speech segment group is determined to indicate no significant divergence. The preset divergence threshold can range from [0.15, 0.35], and is preferably set to 0.25.

[0064] Third, based on the comparison results of the speech segment groups, the target speech segment corresponding to the speech segment group is determined.

[0065] If the comparison results for the corresponding speech segment group show no significant discrepancy, it indicates that the difference between the text after noise reduction and the text after semantic purification is not significant, and there are no errors in noise reduction and semantic purification. In this case, the first speech segment in the speech segment group is directly taken as the target speech segment for that speech segment group.

[0066] If the comparison results for the corresponding speech segment group show significant discrepancies, it indicates a large difference between the denoised and semantically purified text, suggesting potential errors in either the denoising or semantic purification process. Using the first speech segment in the speech segment group as the target speech segment might result in the loss of important textual content. Therefore, the speech segment in the court transcript corresponding to the first speech segment (i.e., the original speech segment in the court transcript that has not undergone denoising and semantic purification and corresponds to the first speech segment) should be used as the target speech segment for the speech segment group. Simultaneously, this target speech segment should be marked as content awaiting review. This embodiment can re-process the target speech segment marked as content to be reviewed by noise reduction and semantic purification to obtain the corresponding first and second speech segments. Then, the re-obtained first and second speech segments are compared. If the comparison result indicates no significant discrepancy, the target speech segment marked as content to be reviewed is replaced with the re-obtained first speech segment, and the mark indicating content to be reviewed is removed. If the comparison result still indicates a significant discrepancy, the target speech segment marked as content to be reviewed is re-processed by noise reduction and semantic purification until the comparison result indicates no significant discrepancy. When the number of times the target speech segment is re-processed by noise reduction and semantic purification reaches a preset threshold, the target speech segment is no longer re-processed by noise reduction and semantic purification. Instead, the target speech segment is output and manually processed by noise reduction and semantic purification. The manually processed text is obtained, and the target speech segment marked as content to be reviewed is replaced with the manually processed text, and the mark indicating content to be reviewed is removed.

[0067] Fourth, the target speech segments corresponding to the speech segment groups are combined in the order of the speech segment groups to obtain the initial standardized text corresponding to the court transcript.

[0068] After identifying the target speech segments corresponding to each speech segment group, the target speech segments need to be sorted according to the order of the speech segment groups. Then, the sorted target speech segments are combined (e.g., text splicing) to obtain the initial regular text corresponding to the court transcript.

[0069] S103. Standardize the initial regularized text to obtain the standard court hearing text.

[0070] After determining the initial standardized text, this embodiment requires standardization processing to obtain the standardized court hearing text. This standardization process includes: legal formalization and / or fact chain reconstruction. Legal formalization refers to replacing content in the initial standardized text with legally formalized text of the same meaning. Examples include: terminology replacement (e.g., replacing "I" with "plaintiff," "you" with "defendant"), applying trial stage templates (e.g., using a template for the court investigation stage, the court debate stage, the final statement stage, and the sentencing stage), and standardizing legal terminology (e.g., replacing "owes me money" with "the defendant owes the plaintiff money"). Fact chain reconstruction refers to reconstructing the speech segments of the initial regularized text according to the fact chain. For example, according to the event chronology chain in the fact chain, several speech segments involving a certain event are sorted by time; according to the causal relationship chain in the fact chain, speech segments with causal relationship are merged; and according to the fact logic chain in the fact chain, speech segments with legal logic are merged according to legal logic, etc.

[0071] In this embodiment, if the standardization process only includes legal writing, the initial standardized text will be legally written, and the resulting court hearing written text will be used as the court hearing standard text. If the standardization process only includes fact chain reconstruction, the initial standardized text will be fact chain reconstructed, and the resulting fact chain reconstructed text will be used as the court hearing standard text. If the standardization process includes both legal writing and fact chain reconstruction, the initial rule text will be legally written to obtain the court hearing written text, and then the court hearing written text will be fact chain reconstructed to obtain the fact chain reconstructed text, which will be used as the court hearing standard text. Alternatively, the initial rule text will be fact chain reconstructed to obtain the fact chain reconstructed text, and then the fact chain reconstructed text will be legally written to obtain the court hearing written text, which will be used as the court hearing standard text.

[0072] In this embodiment, the replacement of titles in the legal writing process can be achieved by pre-constructing a title mapping table corresponding to each role tag. The written title corresponding to the title in the initial standardized text can be queried from the title mapping table corresponding to the role tag of the speech segment to which the title belongs in the initial standardized text, and the title in the initial standardized text can be replaced with the written title corresponding to that title. The standardization of legal terminology in the legal writing process can be achieved by pre-constructing a legal terminology database containing several legal terms corresponding to spoken language. Each statement in the initial standardized text is matched and analyzed with each spoken language in the legal terminology database, and the statement in the initial standardized text that matches the spoken language in the legal terminology database is replaced with the legal term corresponding to that spoken language in the legal terminology database. The application of trial stage templates in the legal writing process involves first performing semantic analysis on the initial standardized text, extracting the content of each trial stage from the initial standardized text, and then applying a pre-set trial stage template to the content of each trial stage. In addition, legal text processing can also employ a large language model, with prompts including address replacement, court template application, and legal terminology standardization. The initial standardized text is input into the large language model, which then performs legal text processing based on the prompts, outputting the court text. For example, the prompts could be set as: "You are a court clerk. Please perform legal text processing on the following text: 1. Replace all colloquial addresses with legal addresses ("I" → "plaintiff", "you" → "defendant"); 2. Identify the current trial stage and apply the corresponding template; 3. Replace colloquial expressions with legal terms ("owes me money" → "overdue payment"). Output the processed text."

[0073] In this embodiment, the fact chain reconstruction process can be performed by training a fact chain reconstruction model. The model is trained using training samples carrying reconstructed text labels, with the training objective being that the predicted reconstructed text output by the model matches the reconstructed text labels. The initial normalized text is input into the fact chain reconstruction model, which outputs the reconstructed fact chain text. This model can be a GNN model, or similar. Alternatively, the fact chain reconstruction process can employ a large language model with set prompts. The initial normalized text is input into the large language model, which performs fact chain reconstruction based on the prompts and outputs the reconstructed fact chain text. For example, the prompts could be set to "Please reconstruct the following text according to the fact chain: 1. Arrange the text in chronological order of events; 2. Merge statements with causal relationships (e.g., 'Because A, therefore B'); 3. Group the text according to legal logic (evidence presentation → cross-examination → debate → determination)."

[0074] S104. Based on the case metadata corresponding to the standard court hearing text and the court hearing transcript, determine the court hearing record corresponding to the court hearing transcript.

[0075] This embodiment requires obtaining case metadata corresponding to the court transcript, such as case number, court time, court location, composition of the collegial panel, and party information. According to the pre-set court transcript template, the standard court text and the case metadata corresponding to the court transcript are combined and spliced ​​to obtain a court transcript that matches the court transcript template, i.e., the court transcript corresponding to the court transcript.

[0076] As described above, the court transcript standardization method proposed in this application involves denoising the transcribed court transcript to obtain a first processed text, semantically refining the transcribed text to obtain a second processed text, determining the initial standardized text corresponding to the transcribed text based on the comparison between the first and second processed texts, standardizing the initial standardized text to obtain a standardized court transcript, and determining the court transcript corresponding to the transcribed text based on the case metadata corresponding to the standardized court transcript and the transcribed text. This embodiment's technical solution enables automatic standardization of transcribed court transcripts, improving the efficiency of transcript production. Furthermore, by decoupling denoising and semantic refining into two independent parallel tracks and comparing the outputs of the two tracks, the initial standardized text is determined based on the comparison results. This avoids the loss of text content due to misjudgment in single-track text processing, thus improving the accuracy of court transcripts.

[0077] As an optional implementation, see [link to implementation details]. Figure 2 As shown, in another embodiment of this application, step S103, which involves standardizing the initial regularized text to obtain the court hearing standardized text, specifically includes the following steps: S201. Based on the pre-constructed knowledge graph of courtroom roles, the initial standardized text is processed into a legally written text to obtain a written courtroom text, which is then used as the standard courtroom text.

[0078] This embodiment can pre-construct a court trial role knowledge graph. The nodes in this court trial role knowledge graph include: roles, legal titles, trial stage templates, and a legal terminology database. Among them, the role nodes can include: judge, presiding judge, judge, plaintiff, defendant, third party, litigation agent, witness, court clerk, prosecutor, defense lawyer, defendant, etc. The legal title nodes can include the mapping between the legal normative title and the colloquial title corresponding to each role. The trial stage template nodes include: court investigation stage template, court debate stage template, final statement stage template, sentencing stage template, etc. The legal terminology database nodes include: mapping between commonly used civil legal terms and colloquial expressions or mapping between commonly used criminal legal terms and colloquial expressions.

[0079] This embodiment, based on a pre-constructed knowledge graph of courtroom roles, performs legal writing standardization processing on the initial standardized text to obtain a written courtroom text, which is then used as the standard courtroom text. The legal writing standardization processing includes: terminology replacement, application of trial stage templates, and standardization of legal terminology.

[0080] In this embodiment, the appellation replacement can be performed by querying the mapping between the legal normative appellation and the colloquial appellation corresponding to the role tag of the speaking segment to which the colloquial appellation belongs in the initial normalized text from the pre-constructed court role knowledge graph, and then querying the legal normative appellation corresponding to the colloquial appellation from the mapping, and replacing the colloquial appellation in the initial normalized text with the corresponding legal normative appellation; the legal terminology standardization can be performed by querying the mapping between common legal terms and colloquial expressions in the legal terminology library node in the pre-constructed court role knowledge graph, and then querying the legal terminology corresponding to the colloquial expression in the initial normalized text, and replacing the colloquial expression in the initial normalized text with the corresponding legal terminology; the court stage template application first performs semantic analysis on the initial normalized text, extracts the content of each court stage from the initial normalized text, obtains the templates for each court stage from the pre-constructed court role knowledge graph, and applies the content of each court stage to the corresponding court stage template. In addition, legal text processing can also employ a large language model, with prompts including address replacement, court template application, and legal terminology standardization. The initial standardized text is input into the large language model, which then performs legal text processing based on the prompts, outputting the court text. For example, the prompt could be set as: "You are a court clerk. Based on the court role knowledge graph, please perform legal text processing on the following text: 1. Replace all colloquial addresses with legal addresses ("I" → "plaintiff", "you" → "defendant"); 2. Identify the current court stage and apply the corresponding template; 3. Replace colloquial expressions with legal terms ("owes me money" → "overdue payment"). Output the processed text."

[0081] As an optional implementation, see [link to implementation details]. Figure 3 As shown, in another embodiment of this application, step S103, which involves standardizing the initial regularized text to obtain the court hearing standardized text, specifically includes the following steps: S301. Reconstruct the fact chain of the written court hearing according to the pre-set fact chain to obtain the standard court hearing text.

[0082] This embodiment pre-sets a fact chain, which includes at least one of an event chronology chain, a causal relationship chain, and a fact logic chain. The fact logic chain is a judicial logic chain, such as the judicial logic of "presenting evidence → cross-examination → debate → determination".

[0083] This embodiment requires reconstructing the written court transcript according to a pre-set fact chain. Specifically, it involves reconstructing the speech segments based on the fact chain. For example, it may merge speech segments corresponding to events within the same timeframe according to the event chronological chain, merge speech segments with causal relationships according to the causal chain, and merge speech segments with legal logic according to the factual logic chain. Events with timestamp differences within a preset threshold are considered to be within the same timeframe. The preset threshold ranges from 5 to 30 minutes, with a preferred value of 15 minutes.

[0084] In one specific implementation, the written court hearing text is reconstructed according to a pre-set fact chain to obtain the standard court hearing text, which specifically includes the following steps: First, according to the preset extraction elements, the written text of the court hearing is extracted to obtain the text elements in the written text of the court hearing.

[0085] This embodiment pre-sets preset extraction elements, specifically including at least one of time, place, person, event, claim, evidence, cross-examination, debate, and conclusion. Based on these preset extraction elements, elements are extracted from each speech segment in the written court transcript, resulting in text elements for each speech segment. In this embodiment, element extraction can be rule-based, pre-setting regular expressions or keyword rules corresponding to each preset extraction element, and extracting text elements from the written court transcript based on these rules. Alternatively, it can employ a pre-trained element extraction model. This model is trained using sample text carrying text element labels, with the training objective being that the predicted text elements output by the model match the text element labels. The written court transcript is then input into the trained element extraction model to obtain text elements. The element extraction model can be an NLP model, or a large language processing model. The model is configured with prompts for feature extraction. The written court text is input into the large language model, which then extracts features based on these prompts and outputs the text elements from the written court text. For example, the prompt could be set to: "Please extract the following elements from the following written court text and output them in JSON format: {Time, Location, Person:[{Role, Name}], Event, Claim:[{Subject, Content}], Evidence:[{Number, Name, Submitter}], Cross-examination:[{Crossing Party, Object, Opinion}], Debate:[{Subject, Content}], Conclusion}. Unmentioned fields should be filled with null."

[0086] Second, based on the text elements and the pre-set chain of facts, the speech segments in the written court text are reconstructed to obtain at least one reconstructed paragraph, and all the reconstructed paragraphs are combined to obtain the standard court text.

[0087] This embodiment reconstructs the spoken segments in the written court hearing text according to the textual elements and pre-set factual chains, obtaining at least one reconstructed segment. All reconstructed segments are then combined to form the standard court hearing text. For example, spoken segments corresponding to events within the same timeframe are merged according to the event chronological chain in the factual chain; spoken segments with causal relationships are merged according to the causal chain in the factual chain; and spoken segments with legal logic are merged according to the factual logic chain in the factual chain.

[0088] In this embodiment, the fact chain reconstruction can be performed by training a fact chain reconstruction model. The model is trained using training samples carrying reconstructed text labels, with the training objective being that the predicted reconstructed text output by the model matches the reconstructed text labels. The written court text is input into the fact chain reconstruction model, which outputs at least one reconstructed paragraph. The fact chain reconstruction model can employ a GNN model, etc. Alternatively, the fact chain reconstruction process can utilize a large language model, with prompts set for the process. The written court text is input into the large language model, which performs fact chain reconstruction based on the prompts, outputting at least one reconstructed paragraph. For example, the prompts could be set to "Please reconstruct the following text according to the fact chain: 1. First, merge the speech segments corresponding to events in the same time period according to the time of occurrence; 2. Merge statements with causal relationships (e.g., "Because A, therefore B"); 3. Group according to judicial logic (presentation of evidence → cross-examination → debate → determination)."

[0089] As an optional implementation, see [link to implementation details]. Figure 4 As shown, in another embodiment of this application, step S104, determining the court transcript corresponding to the court transcript based on the case metadata corresponding to the court transcript and the court transcript, specifically includes the following steps: S401. According to the pre-set court transcript template, combine the case metadata and court transcript text corresponding to the court transcript to obtain the initial transcript.

[0090] This embodiment requires obtaining case metadata corresponding to the court transcript, such as case number, court time, court location, composition of the collegial panel, and party information. According to the pre-set court transcript template, the standard court transcript and the case metadata corresponding to the court transcript are combined and spliced ​​to obtain an initial transcript that matches the court transcript template.

[0091] S402. Based on the pre-set standardized court transcript template library, the initial transcript is verified. If the verification passes, the initial transcript is used as the court transcript corresponding to the court transcript of the court transcript transcription text.

[0092] This embodiment pre-sets a standardized court transcript template library, which includes standardized templates for civil litigation court transcripts or criminal litigation court transcripts. The appropriate template library (civil litigation or criminal litigation) is determined based on the nature of the current trial case. In this embodiment, the standardized court transcript template library may contain templates for at least one trial stage.

[0093] The initial transcript undergoes formal verification based on a standardized court transcript template library. If the verification passes, the initial transcript is used as the court transcript corresponding to the transcribed court transcript. During the formal verification, the initial transcript is matched against various templates in the standardized court transcript template library. If the initial transcript matches any template in the library, the verification is considered successful; if it does not match at least one template, the verification fails. In this embodiment, the matching analysis between the initial transcript and templates in the standardized court transcript template library can be performed using similarity calculation. If the similarity between the template and the content of the initial transcript reaches a preset threshold, the template is considered a match; otherwise, it is considered a mismatch.

[0094] As an optional implementation, see [link to implementation details]. Figure 5 As shown, in another embodiment of this application, step S104, determining the court transcript corresponding to the court transcript based on the case metadata corresponding to the court transcript and the court transcript, specifically includes the following steps: S501. If the verification fails, the speech segments that failed the verification in the initial transcript are marked as content to be reviewed, and the error type of the content to be reviewed is determined.

[0095] If the verification fails, the template in the court transcript standard database that failed the verification with the initial transcript is identified as the verification failed template. Based on the verification failed template, the speech segments in the initial transcript that failed the verification are identified. That is, the speech segments in the initial transcript that need to be matched and analyzed with the verification failed template are identified as the verification failed speech segments. For example, if the verification failed template is a court investigation stage template, then the speech segments in the initial transcript that belong to the court investigation stage are identified as the verification failed speech segments.

[0096] This embodiment marks the speech segments that failed validation in the initial transcript as content to be reviewed and determines the error type of the content to be reviewed. This embodiment determines the error type based on the mismatch between the validation-failed template corresponding to the content to be reviewed and the initial transcript. For example, if the content to be reviewed lacks a required field from the template (e.g., "Plaintiff's Statement" is a required field in the template during the court investigation stage, but the content to be reviewed does not contain "Plaintiff's Statement"), then the error type is determined to be missing element. If the content to be reviewed contains titles that do not conform to the legal norms bound by the template (e.g., "I believe" instead of "The plaintiff believes"), then... If the error type is determined to be a title error; if the statements in the content to be reviewed do not conform to the legal terminology norms constrained by the template (e.g., "he owes me money" instead of "the defendant owes the plaintiff money"), then the error type is determined to be non-standard legal terminology; if the content to be reviewed contains content that does not exist in the template for this stage but exists in templates for other stages, then the error type is determined to be a template application error in the trial stage; if the sequence of events in the content to be reviewed does not conform to the sequence constrained by the template, then the error type is determined to be an event chronology error; if the content to be reviewed is self-contradictory (e.g., the claims do not correspond to the evidence), then the error type is determined to be a break in the factual logical chain.

[0097] In this embodiment, determining the error type of the content to be reviewed can be done using a large language model. A prompt instruction for determining the error type is set, and the content to be reviewed is input into the large language model. The large language model analyzes the error type based on the prompt instruction and outputs the corresponding error type. For example, the prompt instruction could be set as: "You are a court transcript verifier. The following is the standard template for the 'Court Investigation' stage: {Template Content}. The following is the speech segment to be verified: {Content to be Reviewed}. Please determine whether the speech segment meets the template requirements. If it does, output 'Pass'; if it does not, output 'Fail' and select from the following error types: {Error Type List E1-E10}, and explain the reason." Here, the error type list E1-E10 is a pre-set error type list. This embodiment can also use a multi-agent verification method. Corresponding agents are set for various error types, each agent is responsible for verifying one type of error, multiple agents execute in parallel, and finally, the verification results of multiple agents are summarized to obtain the final error type. Different error types can be validated using different agent types. For example, for an error type of missing element, the agent can use JSON Schema validation; for an error type of broken fact logic chain, the agent can use large language model inference validation. This embodiment can also employ a rule-based approach, setting a corresponding rule for each error type and executing the rules using a rule engine to determine the error type of the content to be reviewed.

[0098] S502. Based on the error type, the content to be reviewed is reorganized to obtain the reorganized content of the content to be reviewed.

[0099] In this embodiment, after determining the error type corresponding to the content to be reviewed, the content to be reviewed is re-entered into the step corresponding to the error type for reprocessing, thereby obtaining the reprocessed content as the reorganized content of the content to be reviewed. For example, if the error type is missing elements, the step corresponding to this error type is determined to be "denoising the court transcript to obtain the first processed text, and semantically refining the court transcript to obtain the second processed text." The content to be reviewed is then re-entered into this step, denoising is performed to obtain the first processed text corresponding to the content to be reviewed, and semantic refining is performed to obtain the second processed text corresponding to the content to be reviewed. Then, based on the comparison between the first processed text and the second processed text, the reorganized content corresponding to the content to be reviewed is determined. If the error type is a title error or non-standard legal terminology, the corresponding step is determined to be "based on the pre-constructed courtroom role knowledge graph, perform legal writing standardization processing on the initial standardized text to obtain the courtroom written text." Then, the content to be reviewed is resubmitted to this step, and the content to be reviewed is subjected to legal writing standardization processing to obtain the re-standardized content corresponding to the content to be reviewed. If the error type is an event chronology error or a broken factual logic chain, the corresponding step is determined to be "according to the pre-set factual chain, reconstruct the factual chain of the courtroom written text to obtain the standardized courtroom text." Then, the content to be reviewed is resubmitted to this step, and according to the pre-set factual chain, the factual chain of the content to be reviewed is reconstructed to obtain the re-standardized content corresponding to the content to be reviewed.

[0100] This embodiment also requires verification of the reorganized content corresponding to the content to be reviewed. If the verification still fails, the error type of the content to be reviewed is determined, and the content to be reviewed is reorganized according to the error type until the reorganized content that can pass the verification is obtained.

[0101] Furthermore, this embodiment can also preset a threshold for the number of re-regulation times. If the number of times the content to be reviewed is re-regulated reaches the threshold, the re-regulation of the content to be reviewed is stopped, and the content to be reviewed and the error type are output so that the content to be reviewed can be manually re-regulated, and the feedback of the manually re-regulated content can be obtained. The manually re-regulated content is then used as the re-regulation content corresponding to the content to be reviewed.

[0102] S503. Based on the reorganized content and the initial transcript, determine the court transcript corresponding to the transcribed court transcript.

[0103] In this embodiment, after determining the reorganized content corresponding to the content to be reviewed, the content to be reviewed in the initial transcript is replaced with the reorganized content corresponding to the content to be reviewed. The replaced transcript is then used as the court transcript corresponding to the court transcript in the court transcript transcription. This court transcript is exported according to a preset output format for easy retrieval. The preset output format can be a Word standard format or a PDF standard format, etc.

[0104] As an optional implementation, another embodiment of this application discloses that this embodiment can pre-train a large language model to complete the entire court transcript standardization process. Specifically, the court transcript is input into the large language model, which performs noise reduction on the transcript to obtain a first processed text, and then performs semantic purification to obtain a second processed text. Based on the comparison between the first and second processed texts, the initial standardized text corresponding to the transcript is determined. Based on a pre-constructed court role knowledge graph, the initial standardized text undergoes legal writing standardization processing to obtain a written court transcript. Following a pre-set fact chain, the written court transcript is then... The process involves reconstructing the fact chain from the transcribed text to obtain the standard court transcript. Following a pre-set court transcript template, the case metadata corresponding to the transcribed text is combined with the standard court transcript to obtain the initial transcript. Based on a pre-set standard court transcript template library, the initial transcript is validated. If the validation passes, the initial transcript is used as the court transcript corresponding to the transcribed text. If the validation fails, the unvalidated segments in the initial transcript are marked as content awaiting review, and the error type of the content awaiting review is determined. Based on the error type, the content awaiting review is reorganized to obtain reorganized content. Based on the reorganized content and the initial transcript, the court transcript corresponding to the transcribed text is determined. For each step of the large language model, corresponding prompts can be set, and each step is executed based on these prompts. Training samples can also be collected for each step, and each step can be trained separately to improve the execution accuracy of each step.

[0105] Exemplary device Accordingly, this application also provides a court transcript formatting device, see [link to relevant documentation]. Figure 6 As shown, the device includes: The text processing module 100 is used to perform noise reduction processing on the transcribed court hearing text to obtain the first processed text, and to perform semantic purification on the transcribed court hearing text to obtain the second processed text. The text determination module 110 is used to determine the initial normalized text corresponding to the court transcript text based on the comparison result of the first processed text and the second processed text. The standardization processing module 120 is used to standardize the initial standardized text to obtain the standardized court hearing text; The transcript determination module 130 is used to determine the court transcript corresponding to the court transcript based on the case metadata corresponding to the court transcript and the court transcript.

[0106] As can be seen from the above description, the court transcript standardization device proposed in this application can automatically standardize the transcribed court transcript, improve the efficiency of court transcript production, and decouple the noise reduction and semantic purification of the transcribed court transcript into two independent tracks for parallel processing. The outputs of the two tracks are then compared, and the initial standardized text is determined based on the comparison results. This can avoid the loss of text content due to misjudgment in single-track text processing, thereby improving the accuracy of court transcripts.

[0107] As an optional implementation, another embodiment of this application discloses a text processing module 100, specifically used for: The court transcript is subjected to spoken noise reduction, and / or, transcription error correction, and / or, text compression to obtain the first processed text.

[0108] As an optional implementation, another embodiment of this application discloses a text processing module 100, which is further used for: According to the pre-set semantic purification content, the court transcript is semantically purified to obtain a second processed text. The semantically purified content includes at least one of the following: core claims, evidence, cross-examination opinions, and key points of defense.

[0109] As an optional implementation, another embodiment of this application discloses a text determination module 110, which includes: an alignment unit, a comparison unit, a target determination unit, and a combination unit.

[0110] An alignment unit is used to align the speech segments of the first processed text and the second processed text to obtain at least one group of speech segments; the group of speech segments includes the first speech segment in the first processed text and the second speech segment in the second processed text, and the first speech segment and the second speech segment are mutually aligned speech segments; The comparison unit is used to compare the first and second speech segments in the speech segment group to obtain the comparison results for the speech segment group. The target determination unit is used to determine the target speech segment corresponding to the speech segment group based on the comparison results of the speech segment group. The combination unit is used to combine the target speech segments corresponding to the speech segment group according to the order of the speech segment group to obtain the initial normalized text corresponding to the court transcript.

[0111] As an optional implementation, another embodiment of this application discloses a comparison unit, specifically used for: Calculate the degree of divergence between the first and second speech segments in the speech segment group based on the edit distance and / or semantic similarity between the first and second speech segments; If the degree of disagreement is greater than the preset degree of disagreement threshold, then the comparison results corresponding to the speech segment group indicate that there is a significant disagreement. If the degree of disagreement is not greater than the preset degree of disagreement threshold, then the comparison result corresponding to the speech segment group indicates that there is no significant disagreement.

[0112] As an optional implementation, another embodiment of this application discloses a template determining unit, specifically used for: If the comparison results of the corresponding speech segment group indicate a significant discrepancy, then the speech segment in the court transcript that corresponds to the first speech segment will be taken as the target speech segment of the speech segment group, and the target speech segment will be marked as content to be reviewed. If the comparison results for the corresponding speech segment group indicate no significant disagreement, then the first speech segment is taken as the target speech segment for the speech segment group.

[0113] As an optional implementation, another embodiment of this application discloses a specification processing module 120, which includes a written specification unit.

[0114] The written standardization unit is used to perform legal written standardization processing on the initial standardized text based on the pre-constructed court role knowledge graph, resulting in a written court text, which is then used as the court standard text. The legal written standardization processing includes: title replacement, court stage template application, and legal terminology standardization. The nodes in the court role knowledge graph include: roles, legal titles, court stage templates, and legal terminology database.

[0115] As an optional implementation, another embodiment of this application discloses that the specification processing module 120 further includes a fact chain reconstruction unit.

[0116] The fact chain reconstruction unit is used to reconstruct the fact chain of the written court hearing according to the pre-set fact chain to obtain the standard court hearing text; wherein, the fact chain includes at least one of the following: event chronology chain, causal relationship chain and fact logic chain.

[0117] As an optional implementation, another embodiment of this application discloses a fact chain reconstruction unit, specifically used for: According to the preset extraction elements, the written text of the court hearing is extracted to obtain the text elements in the written text of the court hearing; the preset extraction elements include at least one of the following: time, place, people, events, claims, evidence, cross-examination, debate, and conclusion; Based on textual elements and pre-set factual chains, the speech segments in the written court trial text are reconstructed to obtain at least one reconstructed paragraph. All reconstructed paragraphs are then combined to obtain the standard court trial text.

[0118] As an optional implementation, another embodiment of this application discloses a transcript determination module 130, specifically used for: According to the pre-set court transcript template, the case metadata corresponding to the court transcript text and the court transcript text are combined to obtain the initial transcript; Based on a pre-set library of standardized court transcript templates, the initial transcript is verified. If the verification passes, the initial transcript is used as the court transcript corresponding to the transcribed court transcript.

[0119] As an optional implementation, another embodiment of this application discloses a transcript determination module 130, which is further used for: If the verification fails, the speech segments that failed the verification in the initial transcript will be marked as content to be reviewed, and the error type of the content to be reviewed will be determined. Based on the error type, the content to be reviewed is reorganized to obtain the reorganized content to be reviewed. Based on the reorganized content and the initial transcript, the court transcript corresponding to the transcribed court transcript was determined.

[0120] The court transcript organizing device provided in this embodiment belongs to the same concept as the court transcript organizing method provided in the above embodiments of this application. It can execute the court transcript organizing method provided in any of the above embodiments of this application and has the corresponding functional modules and beneficial effects for executing the court transcript organizing method. Technical details not described in detail in this embodiment can be found in the specific processing content of the court transcript organizing method provided in the above embodiments of this application, and will not be repeated here.

[0121] Exemplary System Optionally, embodiments of this application also provide a court hearing transcript generation system, see [link to relevant documentation]. Figure 7 As shown, the court transcript generation system includes a court transcript transcription system 200 and a court transcript formatting system 210. Data transmission is possible between the court transcript transcription system 200 and the court transcript formatting system 210. The court transcript transcription system 200 transcribes the speech during the court proceedings to obtain a transcribed text, and then transmits the transcribed text to the court transcript formatting system 210. The court transcript formatting system 210 executes the court transcript formatting method described in any of the above embodiments.

[0122] Exemplary electronic devices Another embodiment of this application also provides an electronic device, see [link to relevant documentation] Figure 8 As shown, the device includes: Memory 300 and processor 310; The memory 300 is connected to the processor 310 and is used to store programs; The processor 310 is used to implement the court transcript organization method disclosed in any of the above embodiments by running the program stored in the memory 300.

[0123] Specifically, the aforementioned electronic device may also include: a bus, a communication interface 320, an input device 330, and an output device 340.

[0124] The processor 310, memory 300, communication interface 320, input device 330, and output device 340 are interconnected via a bus. Among them: A bus can include a pathway for transmitting information between various components of a computer system.

[0125] The processor 310 can be a general-purpose processor, such as a general-purpose central processing unit (CPU), a microprocessor, etc., or an application-specific integrated circuit (ASIC), or one or more integrated circuits used to control the execution of the program of the present invention. It can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), an off-the-shelf programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0126] Processor 310 may include a main processor, as well as a baseband chip, modem, etc.

[0127] The memory 300 stores a program that executes the technical solution of this invention, and may also store an operating system and other key business functions. Specifically, the program may include program code, which includes computer operation instructions. More specifically, the memory 300 may include read-only memory (ROM), other types of static storage devices capable of storing static information and instructions, random access memory (RAM), other types of dynamic storage devices capable of storing information and instructions, disk storage, flash memory, etc.

[0128] Input device 330 may include a device for receiving user input data and information, such as a keyboard, mouse, camera, scanner, light pen, voice input device, touch screen, pedometer, or gravity sensor.

[0129] Output device 340 may include devices that allow information to be output to a user, such as a display screen, printer, speaker, etc.

[0130] The communication interface 320 may include a device that uses any transceiver to communicate with other devices or communication networks, such as Ethernet, Radio Access Network (RAN), Wireless Local Area Network (WLAN), etc.

[0131] The processor 310 executes the program stored in the memory 300 and calls other devices, which can be used to implement any of the steps of the court transcript preparation method provided in the above embodiments of this application.

[0132] Exemplary computer program products and storage media In addition to the methods and devices described above, embodiments of this application may also be computer program products, which include computer program instructions that, when executed by a processor, cause the processor to perform the steps in the court transcript preparation methods according to various embodiments of this application described in the "Exemplary Methods" section of this specification.

[0133] The computer program product can be written in any combination of one or more programming languages ​​to perform the operations of the embodiments of this application. The programming languages ​​include object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computing device, partially on the user's computing device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0134] Furthermore, embodiments of this application may also be storage media storing a computer program, which is executed by a processor in the court transcript preparation method according to various embodiments of this application described in the "Exemplary Methods" section above.

[0135] For the foregoing method embodiments, in order to simplify the description, they are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, because according to this application, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0136] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For apparatus embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0137] The steps in the methods of the various embodiments of this application can be adjusted, merged, or deleted in order according to actual needs, and the technical features described in each embodiment can be replaced or combined.

[0138] The modules and sub-modules in the various embodiments of the present application's devices and terminals can be merged, divided, and deleted according to actual needs.

[0139] It should be understood that the disclosed terminals, devices, and methods can be implemented in other ways, given the several embodiments provided in this application. For example, the terminal embodiments described above are merely illustrative. For instance, the division of modules or sub-modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple sub-modules or modules may be combined or integrated into another module, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or modules, and may be electrical, mechanical, or other forms.

[0140] The modules or submodules described as separate components may or may not be physically separate. The components that constitute a module or submodule may or may not be physical modules or submodules; that is, they may be located in one place or distributed across multiple network modules or submodules. Some or all of the modules or submodules can be selected to achieve the purpose of this embodiment's solution, depending on actual needs.

[0141] Furthermore, the functional modules or sub-modules in the various embodiments of this application can be integrated into one processing module, or each module or sub-module can exist physically separately, or two or more modules or sub-modules can be integrated into one module. The integrated modules or sub-modules described above can be implemented in hardware or in the form of software functional modules or sub-modules.

[0142] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0143] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software unit executed by a processor, or a combination of both. The software unit can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0144] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0145] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for organizing court hearing transcripts, characterized in that, include: The court transcript was denoised to obtain the first processed text, and the court transcript was semantically purified to obtain the second processed text. Based on the comparison results between the first processed text and the second processed text, the initial normalized text corresponding to the court transcript is determined; The initial standardized text is then processed to obtain the standardized court hearing text; Based on the case metadata corresponding to the court hearing transcript and the court hearing transcription text, the court hearing record corresponding to the court hearing transcription text is determined.

2. The method for organizing court transcripts according to claim 1, characterized in that, The noise reduction process performed on the transcribed court transcript yields a first processed text, including: The court transcript is subjected to spoken noise reduction, and / or, transcription error correction is performed on the court transcript, and / or, text compression is performed on the court transcript to obtain the first processed text.

3. The method for organizing court hearing transcripts according to claim 1, characterized in that, The semantic purification of the transcribed court transcript to obtain the second processed text includes: According to the pre-set semantic purification content, the court transcript is semantically purified to obtain a second processed text. The semantically purified content includes at least one of the following: core claims, evidence, cross-examination opinions, and key points of defense.

4. The method for organizing court transcripts according to claim 1, characterized in that, Based on the comparison results between the first processed text and the second processed text, the initial standardized text corresponding to the court transcript is determined, including: The first processed text and the second processed text are aligned to obtain at least one group of speech segments; the group of speech segments includes a first speech segment in the first processed text and a second speech segment in the second processed text, and the first speech segment and the second speech segment are mutually aligned speech segments; The first and second speech segments in the speech segment group are compared to obtain the comparison result corresponding to the speech segment group; Based on the comparison results corresponding to the speech segment group, the target speech segment corresponding to the speech segment group is determined; The target speech segments corresponding to the speech segment groups are combined in the order of the speech segment groups to obtain the initial regularized text corresponding to the court transcript.

5. The method for organizing court transcripts according to claim 4, characterized in that, The first and second speech segments in the speech segment group are compared to obtain the comparison results corresponding to the speech segment group, including: The degree of divergence between the first and second speech segments is calculated based on the edit distance and / or semantic similarity between the first and second speech segments in the speech segment group; If the degree of divergence is greater than a preset degree of divergence threshold, then the comparison result corresponding to the speech segment group indicates that there is a significant divergence. If the degree of divergence is not greater than a preset degree of divergence threshold, then the comparison result corresponding to the speech segment group indicates that there is no significant divergence.

6. The method for organizing court transcripts according to claim 4, characterized in that, Based on the comparison results corresponding to the speech segment group, the target speech segment corresponding to the speech segment group is determined, including: If the comparison results corresponding to the speech segment group indicate a significant discrepancy, then the speech segment in the court transcript that corresponds to the first speech segment is taken as the target speech segment corresponding to the speech segment group, and the target speech segment is marked as content to be reviewed; If the comparison results corresponding to the speech segment group indicate that there is no significant disagreement, then the first speech segment is taken as the target speech segment corresponding to the speech segment group.

7. The method for organizing court hearing transcripts according to claim 1, characterized in that, The initial standardized text is then normalized to obtain the court hearing standard text, including: Based on a pre-constructed knowledge graph of courtroom roles, the initial standardized text is subjected to legal writing standardization processing to obtain a written courtroom text, which is then used as the standard courtroom text. The legal writing standardization processing includes: title replacement, application of courtroom stage templates, and standardization of legal terminology. The nodes in the knowledge graph of courtroom roles include: roles, legal titles, courtroom stage templates, and a legal terminology database.

8. The method for organizing court transcripts according to claim 7, characterized in that, The initial standardized text is then normalized to obtain the standardized court hearing text, which also includes: The court hearing text is reconstructed according to a pre-set fact chain to obtain a standard court hearing text; wherein the fact chain includes at least one of the following: an event chronology chain, a causal relationship chain, and a fact logic chain.

9. The method for organizing court transcripts according to claim 8, characterized in that, Based on a pre-set chain of facts, the written court hearing text is reconstructed to obtain a standard court hearing text, including: According to preset extraction elements, the written text of the court hearing is subjected to element extraction to obtain the text elements in the written text of the court hearing; the preset extraction elements include at least one of the following: time, place, people, events, claims, evidence, cross-examination, debate, and conclusion; Based on the text elements and the pre-set fact chain, the speech segments in the written court text are reconstructed to obtain at least one reconstructed paragraph. All reconstructed paragraphs are then combined to obtain the standard court text.

10. The method for organizing court hearing transcripts according to claim 1, characterized in that, Based on the case metadata corresponding to the court hearing transcript and the court hearing transcription text, the court hearing record corresponding to the court hearing transcription text is determined, including: According to the pre-set court transcript template, the case metadata corresponding to the transcribed court transcript and the standard court transcript are combined to obtain the initial transcript; Based on a pre-set library of standardized court transcript templates, the initial transcript is verified. If the verification passes, the initial transcript is used as the court transcript corresponding to the transcribed court transcript.

11. The method for organizing court transcripts according to claim 10, characterized in that, Based on the case metadata corresponding to the court hearing transcript and the court hearing transcription text, determining the court hearing record corresponding to the court hearing transcription text also includes: If the verification fails, the speech segments that fail the verification in the initial transcript are marked as content to be reviewed, and the error type of the content to be reviewed is determined. Based on the error type, the content to be reviewed is reorganized to obtain the reorganized content of the content to be reviewed; Based on the reorganized content and the initial transcript, the court transcript corresponding to the transcribed court transcript is determined.

12. A court hearing transcript generation system, characterized in that, include: Court hearing transcription system and court hearing transcript organization system; The court hearing transcription system is used to transcribe the speech during the court hearing to obtain the transcribed text, and then transmit the transcribed text to the court hearing transcript system. The court transcript formatting system is used to execute the court transcript formatting method according to any one of claims 1 to 11.

13. An electronic device, characterized in that, include: Memory and processor; The memory is connected to the processor and is used to store programs; The processor is configured to implement the court transcript standardization method as described in any one of claims 1 to 11 by running a program in the memory.

14. A computer program product, characterized in that, It includes computer program instructions that, when executed by a processor, cause the processor to implement the court transcript formatting method as described in any one of claims 1 to 11.

Images