Key information extraction method and device and related equipment
By segmenting and calculating semantic similarity of voice data, the contradiction between real-time performance and high accuracy in traditional technologies is resolved, enabling efficient and accurate extraction of key information in judicial and administrative scenarios, reducing the false negative rate and the workload of manual review.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING TAIXIN TIANCHENG TECHNOLOGY CO LTD
- Filing Date
- 2026-01-08
- Publication Date
- 2026-04-28
AI Technical Summary
Traditional speech structure extraction technology struggles to simultaneously meet the demands of real-time processing and high accuracy in judicial and administrative settings, resulting in delays in acquiring critical information or high false positive rates, thus failing to meet the needs of real-time interrogation and on-site verification.
By receiving voice data in real time and converting it into text segments, combining a legal terminology hot word database and a lightweight extraction model to extract entity slots, and using a large language model to calculate semantic similarity, the semantic scores of candidate and selected entity slots are determined to ensure the accuracy of key information.
It improves the accuracy of key information extraction, reduces the false negative rate, meets the real-time and high-precision requirements in judicial and administrative scenarios, and reduces the workload of manual review.
Smart Images
Figure CN121938355A_ABST
Abstract
Description
Technical Field
[0001] The embodiments of the present invention relate to the field of information processing technology, and in particular to a method, apparatus and related equipment for extracting key information. Background Technology
[0002] In judicial and administrative settings where secure audio evidence is required, audio data serves as a core evidence carrier. The efficiency and accuracy of its structured processing directly impact process efficiency, the legality of evidence preservation, and the accuracy of judicial proceedings. However, traditional audio structure extraction technologies, limited by a single-path processing architecture, struggle to meet the core requirements of both real-time performance and high precision in such scenarios. These limitations manifest as multiple pain points in actual business scenarios, severely hindering improvements in process efficiency. On the one hand, some traditional solutions adopt a pure ASR text-level post-processing mode. Its core logic is to first convert all voice data into text using speech-to-text (ASR) technology, and then extract entity slots and organize information based on the complete text. This mode has particularly prominent drawbacks in scenarios with high real-time requirements, such as judicial interrogations and on-site verification: During the interrogation process, staff need to grasp key information from the statements of the person being investigated (such as the connection of relevant personnel, the time and place of the crime, and the location of the evidence) in real time to adjust the interrogation strategy. However, this mode requires waiting for the complete transcription of the interrogation speech to be completed (usually several hours or even longer, depending on the duration of the speech) before extraction can be carried out, resulting in a significant delay in the acquisition of key information and failing to provide support for real-time interrogation guidance. At the scene of an emergency administrative verification (such as drunk driving investigation or on-site verification of illegal business operations), the investigators need to quickly extract relevant elements from the statements of the parties involved and the testimonies of surrounding witnesses to secure preliminary evidence. However, the post-processing mode after full transcription cannot meet the needs of on-site immediate evidence collection and verification, and is prone to problems such as the omission of key evidence or untimely securing. Furthermore, if transcription errors occur during the full speech-to-text transcription process, the errors will be directly transmitted to the subsequent text extraction stage. Due to the lack of a real-time verification mechanism, the subsequent correction cost is extremely high, further affecting the reliability of the evidence information.
[0003] On the other hand, another type of traditional solution, in pursuit of real-time performance, relies solely on lightweight extraction modes at the frame or fragment level of ASR. This involves real-time parsing of frame-level or fragment-level data from the speech stream to synchronously extract simple entities and output the results. While this mode can achieve low-latency output and meet the immediate information acquisition needs of some scenarios, it suffers from limitations such as the lack of contextual integrity in frame / fragment-level data and the limited extraction capabilities of lightweight models. Consequently, its information extraction accuracy is difficult to guarantee, resulting in high rates of missed and false detections of key entities. In judicial settings, for key entities with standardized formats and strong relevance, such as legal provision numbers, evidence numbers, relevant personnel identity information (e.g., ID card numbers, work units), and amounts, this model is prone to omissions (e.g., missing some characters in long legal provision numbers), false positives (e.g., confusing names of related personnel with similar pronunciations), or fragmented information (e.g., extracting only the surname of the relevant personnel without extracting the complete name) because it cannot rely on complete context for semantic verification. In administrative verification scenarios, the lightweight extraction model struggles to accurately identify core elements such as the legal basis clauses corresponding to violations, the specifications and models of related items, and the subject information of the verified party. This results in the extracted structured information not being directly used as evidence, requiring subsequent manual review and supplementation by staff, which increases the workload. Summary of the Invention
[0004] The embodiments of the present invention provide a method, apparatus and related equipment for extracting key information, which aims to solve the technical problem that traditional technologies are unable to accurately extract key information from the voice data of judicial personnel.
[0005] In a first aspect, embodiments of the present invention provide a method for extracting key information, comprising: Real-time reception of voice data, conversion of the voice data into text segments to obtain multiple text segments, and labeling of all text segments to obtain segment tags for all text segments. The segment tags include at least segment ID, start timestamp and end timestamp. Based on a pre-built legal terminology hot word library and a lightweight extraction model, entity slots are extracted from the voice data and text segments to obtain several candidate entity slots. The candidate entity slots are associated with corresponding candidate tags, and the candidate tags include at least the corresponding segment ID, timestamp, and confidence level. Obtain the complete text corresponding to the voice data. Based on the legal terminology hot word library and the context information of the complete text, extract entity slots and remove duplicates from the complete text to obtain several selected entity slots. The selected entity slots are associated with corresponding selected tags. The selected tags include at least the corresponding text fragment, timeline information and confidence level. Based on the segment IDs and text fragments in the candidate tags and selected tags, the corresponding candidate entity slots and selected entity slots are determined, and the corresponding candidate entity slots and selected entity slots are output to the large language model for semantic similarity calculation. Determine whether the semantic score is greater than or equal to a preset upper limit threshold; If so, the information of the corresponding candidate entity slot is used as the target key information; If not, the information of the corresponding selected entity slot will be used as the target key information.
[0006] Secondly, embodiments of the present invention provide a key information extraction device, comprising: The segmentation module is used to receive voice data in real time, convert the voice data into text segments, obtain multiple text segments, and label all text segments to obtain segment tags for all text segments. The segment tags include at least segment ID, start timestamp, and end timestamp. The first extraction module is used to extract entity slots from the voice data and text segments based on a pre-set legal term hot word library and a lightweight extraction model, to obtain a number of candidate entity slots. The candidate entity slots are associated with corresponding candidate tags, and the candidate tags include at least the corresponding segment ID, timestamp, and confidence level. The second extraction module is used to obtain the complete text corresponding to the voice data. Based on the legal term hot word library and the context information of the complete text, the module extracts entity slots and removes duplicates from the complete text to obtain a number of selected entity slots. The selected entity slots are associated with corresponding selected tags. The selected tags include at least the corresponding text fragments, timeline information and confidence level. The scoring module is used to determine the corresponding candidate entity slots and selected entity slots based on the segment IDs and text fragments in the candidate tags and selected tags, and output the corresponding candidate entity slots and selected entity slots to the large language model for semantic similarity calculation to obtain a semantic score. The judgment module is used to determine whether the semantic score is greater than or equal to a preset upper limit threshold; if so, the information of the corresponding candidate entity slot is used as the target key information; if not, the information of the corresponding selected entity slot is used as the target key information.
[0007] Thirdly, embodiments of the present invention provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the key information extraction method described in the first aspect above.
[0008] Fourthly, embodiments of the present invention provide a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, which, when executed by a processor, causes the processor to perform the key information extraction method described in the first aspect.
[0009] This invention provides a method, apparatus, and related equipment for extracting key information. The method receives voice data in real time, converts the voice data into text segments, obtains multiple text segments, and labels all text segments to obtain segment tags. Based on a pre-built legal terminology hot word library and a lightweight extraction model, entity slots are extracted from the voice data and text segments to obtain several candidate entity slots. The complete text corresponding to the voice data is obtained, and based on the legal terminology hot word library and the context information of the complete text, entity slots are extracted from the complete text and deduplicated to obtain several selected entity slots. The similarity score between the voices of the candidate entity slots and the selected entity slots is calculated to determine the target key information. This method improves the accuracy of information extraction and effectively reduces the missed detection rate of key entities when the voice data is chaotic. Attached Figure Description
[0010] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0011] Figure 1 This is a flowchart illustrating an embodiment of the key information extraction method provided by the present invention. Figure 2 A schematic block diagram of a key information extraction device provided in an embodiment of the present invention. Detailed Implementation
[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0013] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.
[0014] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.
[0015] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.
[0016] Please see Figure 1 This is a flowchart illustrating a key information extraction method provided in an embodiment of the present invention, which includes steps S110 to S150.
[0017] Step S110: Receive voice data in real time, convert the voice data into text segments to obtain multiple text segments, and label all text segments to obtain segment labels for all text segments; In this embodiment, voice data is received in real time, and the voice streaming is converted into text segments using an ASR (Automatic Speech Recognition) system. Basic information such as the segment ID, start timestamp, and end timestamp of each text segment is simultaneously labeled. For example: [0.000s-3.250s] Speaker A: Hello, is this XX customer service center? [3.500s-6.800s] Speaker B: Yes, hello, how can I help you? [7.000s-12.500s] Speaker A: I bought a Bluetooth headset yesterday, but when I used it today, I found that there was no sound in the left ear.
[0018] [13.000s-18.300s] Speaker B: Okay, please don't rush. Could you please provide the order number and the platform on which you made the purchase? [18.500s-22.100s] Speaker A: The order number is 123456789, and it was purchased from the XX official flagship store.
[0019] Step S120: Based on a pre-set legal terminology hot word library and a lightweight extraction model, entity slots are extracted from the voice data and text segments to obtain several candidate entity slots, and the candidate entity slots are associated with corresponding candidate tags. In this embodiment, a pre-built hot word library of legal terminology is constructed, covering common legal provision numbers, evidence types, role names of judicial personnel, and professional legal terms, providing keyword support for subsequent frame-level and text-level extraction. Each audio segment and text segment is simultaneously input into a pre-built lightweight extraction model. The model uses the hot word library to extract entity slots (such as names of judicial personnel, evidence numbers, and case occurrence times), resulting in several candidate entity slots. Each candidate slot is marked as a "draft" and associated with its corresponding segment ID, timestamp, and confidence level. Simultaneously, the model uses a context memory mechanism to cache key information from previous and subsequent segments, optimizing the continuous extraction effect across segments. It is important to note that the candidate entity slots output by the model need to be verified to ensure they conform to specifications, avoiding errors in data format or structure. For example, a correctly formatted entity slot is... The entity slot with incorrect data format is .
[0020] Step S130: Obtain the complete text corresponding to the voice data. Based on the legal terminology hot word library and the context information of the complete text, extract entity slots and remove duplicates from the complete text to obtain several selected entity slots. The selected entity slots are associated with corresponding selected tags. In this embodiment, speech recognition technology is used to obtain the complete text corresponding to the speech data. Based on a legal terminology hot word database and the contextual information of the complete text, entity slots in the complete text are accurately identified. Multiple mentions of the same entity slot are then merged and deduplicated (e.g., different titles for the same person to be verified are standardized to a standard name), resulting in several selected entity slots. Simultaneously, the relationships between these selected entity slots are established (e.g., the attribution relationship between evidence and relevant personnel). Furthermore, to facilitate the distinction between candidate entity slots and selected entity slots, each extracted selected entity slot is marked as "pending confirmation" and associated with the corresponding text fragment, timeline information, and confidence level. It should be noted that the selected entity slots page needs to be verified to ensure compliance with specifications.
[0021] Step S140: Based on the segment ID and text fragment in the candidate tags and selected tags, determine the corresponding candidate entity slots and selected entity slots, and output the corresponding candidate entity slots and selected entity slots to the large language model for semantic similarity calculation to obtain a semantic score; Step S150: Determine whether the semantic score is greater than or equal to a preset upper limit threshold; Step S161: If yes, then use the information of the corresponding candidate entity slot as the target key information; Step S162: If not, then use the information of the corresponding selected entity slot as the target key information.
[0022] In this embodiment, candidate entity slots and corresponding selected entity slots for the same time period are determined based on the timestamps of candidate tags and the timelines of selected tags. A Large Language Model (LLM) is used to calculate the semantic similarity between candidate entity slots and their corresponding selected entity slots of the same type. The semantic similarity calculation methods include, but are not limited to, directly scoring the semantic similarity between candidate entity slots and their corresponding selected entity slots driven by prompt words, or using the embedding capabilities of the LLM to convert candidate entity slots and their corresponding selected entity slots into high-dimensional vectors, and quantifying the semantic similarity by calculating the angle between the vectors using cosine similarity. The semantic score is then determined to be greater than or equal to a preset upper threshold. If the semantic score is greater than or equal to the preset upper threshold, it indicates that the candidate entity slots and their corresponding selected entity slots are highly semantically similar, and the information from the candidate entity slots is directly used as the target key information. If the semantic score is less than the preset upper threshold, it indicates that the semantics of the candidate entity slots and the selected entity slots differ significantly. Since the extraction of selected entity slots is context-dependent and more accurate than that of candidate entity slots, the information from the selected entity slots is used as the target key information. In addition, each entity slot that is confirmed as key target information is marked as "confirmed" to facilitate user differentiation of entity operations.
[0023] In one embodiment, if the semantic score is less than a preset lower threshold, it indicates that the semantics of the candidate entity slot and the corresponding selected entity slot are too different, which may be due to a large error in the extraction of the entity slot. Therefore, the corresponding entity slot is marked as to be reviewed, and the entity slot to be reviewed, along with the corresponding voice data and text segments, are visualized for the user to correct.
[0024] In one embodiment, to ensure information immutability and ease of traceability, complete traceability information is added to each "confirmed" entity slot, including the source voice segment ID, text segment ID, complete timeline (accurate to the frame level), and speaker role (determined through a speaker separation model and role mapping system). A hash calculation is performed on the complete information of each field (including field value, traceability information, and confidence level) to generate a unique hash signature. Specifically, a speaker separation model and role mapping system are pre-built and trained to adapt to judicial scenarios (police officers, individuals under investigation, witnesses, victims, etc.), laying the foundation for subsequent field attribution determination. For example, the entire recording is simulated as a multi-person chat recording. Tools are used to cut the recording into segments, and similar segments are assigned to the same person based on voice characteristics, numbered as spk1, spk2, etc., thus identifying voice data belonging to the same person. The role corresponding to each number is determined by content and tone: for example, those who read aloud or ask questions are often police officers / judges, while those who mainly answer or recount their experiences are often individuals under investigation / witnesses. Finally, the results of "number → role" are written back to each segment to obtain "who said what and when, and which role they belong to." That is, the sound distinguishes the person to whom the speech data belongs, the content determines the corresponding role, and then it is filled back into the segment to complete the mapping from audio to role.
[0025] This method receives speech data in real time, converts it into text segments, obtains multiple text segments, and labels all text segments. Based on a pre-built legal terminology hotspot database and a lightweight extraction model, entity slots are extracted from the speech data and text segments to obtain several candidate entity slots. The complete text corresponding to the speech data is obtained, and based on the legal terminology hotspot database and the context information of the complete text, entity slots are extracted from the complete text and deduplicated to obtain several selected entity slots. The similarity score between the speech of the candidate entity slots and the selected entity slots is calculated to determine the target key information. This method improves the accuracy of information extraction and effectively reduces the missed detection rate of key entities when the speech data is chaotic.
[0026] This invention also provides a key information extraction apparatus, which is used to execute any embodiment of the aforementioned key information extraction method. Specifically, please refer to... Figure 2 , Figure 2 This is a schematic block diagram of a key information extraction device provided in an embodiment of the present invention. The key information extraction device 100 can be configured in a server.
[0027] like Figure 2 As shown, the key information extraction device 100 includes a segmentation module 110, a first extraction module 120, a second extraction module 130, a verification module 140, and a judgment module 150.
[0028] The segmentation module 110 is used to receive voice data in real time, convert the voice data into text segments, obtain multiple text segments, and label all text segments to obtain segmentation tags for all text segments. The segmentation tags include at least a segmentation ID, a start timestamp, and an end timestamp. The first extraction module 120 is used to extract entity slots from the voice data and text segments based on a pre-set legal term hot word library and a lightweight extraction model, to obtain a number of candidate entity slots. The candidate entity slots are associated with corresponding candidate tags, and the candidate tags include at least the corresponding segment ID, timestamp, and confidence level. The second extraction module 130 is used to obtain the complete text corresponding to the voice data, and based on the legal term hot word library and the context information of the complete text, extract entity slots and remove duplicates from the complete text to obtain a number of selected entity slots. The selected entity slots are associated with corresponding selected tags, and the selected tags include at least the corresponding text fragments, timeline information and confidence level. The scoring module 140 is used to determine the corresponding candidate entity slots and selected entity slots based on the segment IDs and text fragments in the candidate tags and selected tags, and output the corresponding candidate entity slots and selected entity slots to the large language model for semantic similarity calculation to obtain a semantic score. The judgment module 150 is used to determine whether the semantic score is greater than or equal to a preset upper limit threshold; if so, the information of the corresponding candidate entity slot is used as the target key information; if not, the information of the corresponding selected entity slot is used as the target key information.
[0029] In one embodiment, the key information extraction device 100 includes, as well as: The visualization module is used to visualize the entity slots to be reviewed, along with the corresponding voice data and text segments, and to display a manual correction window.
[0030] In one embodiment, the key information extraction device 100 includes, as well as: The hash processing module is used to perform hash calculations on the complete information of each field in the key information set to generate a hash signature. The complete information includes at least the field value corresponding to the field, traceability information, and confidence level.
[0031] This invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the key information extraction method described above.
[0032] In another embodiment of the invention, a computer-readable storage medium is provided. This computer-readable storage medium may be a non-volatile computer-readable storage medium. The computer-readable storage medium stores a computer program that, when executed by a processor, causes the processor to perform the key information extraction method as described above.
[0033] Those skilled in the art will readily understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. Those skilled in the art will 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 composition and steps of each example have been generally described in terms of function 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 invention.
[0034] In the embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Units with the same function may be grouped into one unit. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interfaces, devices, or units, or it may be an electrical, mechanical, or other form of connection.
[0035] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of the present invention, depending on actual needs.
[0036] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.
[0037] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), magnetic disks, or optical disks.
[0038] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A method for extracting key information, characterized in that, include: Real-time reception of voice data, conversion of the voice data into text segments to obtain multiple text segments, and labeling of all text segments to obtain segment tags for all text segments. The segment tags include at least segment ID, start timestamp and end timestamp. Based on a pre-built legal terminology hot word library and a lightweight extraction model, entity slots are extracted from the voice data and text segments to obtain several candidate entity slots. The candidate entity slots are associated with corresponding candidate tags, and the candidate tags include at least the corresponding segment ID, timestamp, and confidence level. Obtain the complete text corresponding to the voice data. Based on the legal terminology hot word library and the context information of the complete text, extract entity slots and remove duplicates from the complete text to obtain several selected entity slots. The selected entity slots are associated with corresponding selected tags. The selected tags include at least the corresponding text fragments, timeline information and confidence level. Based on the segment ID and text fragment in the candidate tags and selected tags, the corresponding candidate entity slots and selected entity slots are determined, and the corresponding candidate entity slots and selected entity slots are output to the large language model for semantic similarity calculation to obtain a semantic score. Determine whether the semantic score is greater than or equal to a preset upper limit threshold; If so, the information of the corresponding candidate entity slot is used as the target key information; If not, the information of the corresponding selected entity slot will be used as the target key information.
2. The key information extraction method as described in claim 2, characterized in that, Also includes: Determine whether the semantic score is less than a preset lower threshold. If so, the corresponding entity slot will be marked as pending review, and the entity slot pending review, along with the corresponding voice data and text segments, will be visualized.
3. The key information extraction method as described in claim 1, characterized in that, The legal terminology hot word database contains at least common legal provision numbers, types of evidence, names of judicial roles, and professional legal terms.
4. The key information extraction method as described in claim 1, characterized in that, After performing information completion and structuring processing on the confirmed entity slot to obtain a structured set of key information, the process includes: A hash calculation is performed on the complete information of each field in the key information set to generate a hash signature. The complete information includes at least the field value corresponding to the field, traceability information, and confidence level.
5. The key information extraction method as described in claim 1, characterized in that, include: Verify whether the candidate entity slots and selected entity slots conform to the slot specifications; If it matches, keep it. If it doesn't match, remove it.
6. A key information extraction device, characterized in that, include: The segmentation module is used to receive voice data in real time, convert the voice data into text segments, obtain multiple text segments, and label all text segments to obtain segment tags for all text segments. The segment tags include at least segment ID, start timestamp, and end timestamp. The first extraction module is used to extract entity slots from the voice data and text segments based on a pre-set legal term hot word library and a lightweight extraction model, to obtain a number of candidate entity slots. The candidate entity slots are associated with corresponding candidate tags, and the candidate tags include at least the corresponding segment ID, timestamp, and confidence level. The second extraction module is used to obtain the complete text corresponding to the voice data. Based on the legal term hot word library and the context information of the complete text, the module extracts entity slots and removes duplicates from the complete text to obtain a number of selected entity slots. The selected entity slots are associated with corresponding selected tags. The selected tags include at least the corresponding text fragments, timeline information and confidence level. The scoring module is used to determine the corresponding candidate entity slots and selected entity slots based on the segment IDs and text fragments in the candidate tags and selected tags, and output the corresponding candidate entity slots and selected entity slots to the large language model for semantic similarity calculation to obtain a semantic score. The judgment module is used to determine whether the semantic score is greater than or equal to a preset upper limit threshold; if so, the information of the corresponding candidate entity slot is used as the target key information; if not, the information of the corresponding selected entity slot is used as the target key information.
7. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the key information extraction method as described in any one of claims 1 to 5.
8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, causes the processor to perform the key information extraction method as described in any one of claims 1 to 5.