Information input method and device, computer device and storage medium

Abstract

The application discloses an information entry method and device, computer equipment and a storage medium, relates to the technical field of information processing, and can be applied to the fields of finance and insurance and medical insurance, and comprises the following steps: acquiring recording data; performing noise reduction processing on the recording data to obtain enhanced recording data; sequentially performing speaker segmentation and clustering, speech recognition and speaker role labeling on the enhanced recording data to obtain role-labeled dialogue text; inputting the role-labeled dialogue text into an information extraction model to acquire claim investigation information and fill the claim investigation information into corresponding fields of a corresponding claim case form; the claim investigation information comprises personal information, work information, guardian information and nursing information; wherein, a first address in the personal information and a second address in the work information are determined by complementing incomplete addresses recognized in the role-labeled dialogue text by the information extraction model. According to the scheme, the information distortion problem in the information entry process is reduced, and the accuracy of information entry is improved.

Description

Technical Field

[0001] This invention relates to the field of information processing technology and can be applied to the fields of finance and insurance and medical insurance. In particular, it relates to an information input method, device, computer equipment and storage medium. Background Technology

[0002] During insurance claims visits, claims personnel must travel to the hospital to conduct an on-site visit and gather information about the injured person and their family's financial situation. This step is crucial for determining the amount of compensation and assessing items such as lost wages, nursing care expenses, and living expenses for dependents. Current procedures typically involve claims personnel recording conversations by hand or phone at the visit site, then reviewing the recordings and notes after the visit to manually input the conversations into the claims system.

[0003] However, due to the intertwined speeches of claims personnel and injured persons during the visit, it is difficult for human listeners to quickly identify the speaker of a large amount of colloquial information when listening to the recording. Furthermore, memory bias can easily lead to discrepancies between the entered information and the actual situation, resulting in the problem of distorted entered information. Summary of the Invention

[0004] This invention provides an information input method, apparatus, computer equipment, and storage medium to solve the problem that existing information input methods are prone to causing information distortion.

[0005] Firstly, an information entry method is provided, including: Acquire audio recordings, including conversations between claims adjusters and the target injured party; The audio data is subjected to noise reduction processing to obtain enhanced audio data; The enhanced recording data is sequentially subjected to speaker segmentation and clustering, speech recognition, and speaker role labeling to obtain role-labeled dialogue text. The role-labeled dialogue text is a global turn-taking text sequence with role labels, and the role labels include claims personnel and target injured persons. The character-annotated dialogue text is input into a preset information extraction model. The information extraction model is used to extract information and complete addresses from the character-annotated dialogue text to obtain the claims visitation information of the target injured person output by the information extraction model. The claims visitation information includes personal information, work information, caregiver information, and nursing information. Among them, the personal information includes a first address, and the work information includes a second address. The first address and the second address are determined by the information extraction model after completing the incomplete addresses identified in the character-annotated dialogue text. Fill the corresponding fields of the claim visitation information into the corresponding claim case form.

[0006] Secondly, an information input device is provided, comprising: The acquisition module is used to acquire audio data, which includes dialogue data between the claims adjuster and the target injured party; A noise reduction module is used to perform noise reduction processing on the recording data to obtain enhanced recording data; The differentiation module is used to sequentially perform speaker segmentation and clustering, speech recognition, and speaker role labeling on the enhanced recording data to obtain role-labeled dialogue text. The role-labeled dialogue text is a global turn-taking text sequence with role labels, and the role labels include claims personnel and target injured persons. An extraction module is used to input the character-annotated dialogue text into a preset information extraction model, and use the information extraction model to perform information extraction and address completion processing on the character-annotated dialogue text to obtain the claims visitation information of the target injured person output by the information extraction model. The claims visitation information includes personal information, work information, caregiver information, and nursing information; wherein, the personal information includes a first address, the work information includes a second address, and the first address and the second address are determined by the information extraction model after completing the incomplete addresses identified in the character-annotated dialogue text. The fill module is used to fill the claim visitation information into the corresponding fields of the corresponding claim case form.

[0007] Thirdly, a computer device is provided, 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 steps of the above-described information entry method.

[0008] Fourthly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the steps of the above-described information entry method.

[0009] In the aforementioned information entry method, device, computer equipment, and storage medium, the recorded data of the visitation dialogue is denoised to obtain enhanced recording data. Then, speaker separation, speech recognition, and speaker role labeling are performed on the enhanced recording data to obtain role-labeled dialogue text. After inputting the role-labeled dialogue text into the information extraction model, the model extracts personal information, work information, caregiver information, and nursing information from the target injured person's turns of speech based on the role labels. It also completes any incomplete addresses identified to obtain complete addresses, and finally outputs structured claims visitation information and fills it into the corresponding fields of the corresponding claims form. Because the role-labeled dialogue text clearly identifies the speaker in each segment, the information extraction model can selectively extract information from the target injured person's statements, effectively avoiding the misinterpretation of claims personnel's questions or restatements as the target injured person's statements. Furthermore, the automatic extraction and filling of information replaces the manual input process after listening to the recordings, further effectively avoiding information distortion caused by memory bias and improving the accuracy of information entry. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only 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 schematic diagram of an application environment for an information entry method according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating an information entry method according to an embodiment of the present invention; Figure 3 This is another flowchart illustrating the information entry method in one embodiment of the present invention; Figure 4 This is another flowchart illustrating the information entry method in one embodiment of the present invention; Figure 5 This is another flowchart illustrating the information entry method in one embodiment of the present invention; Figure 6 This is another flowchart illustrating the information entry method in one embodiment of the present invention; Figure 7 This is another flowchart illustrating the information entry method in one embodiment of the present invention; Figure 8 This is another flowchart illustrating the information entry method in one embodiment of the present invention; Figure 9 This is a schematic diagram of the structure of an information input device in one embodiment of the present invention; Figure 10 This is a schematic diagram of the structure of a computer device according to 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] The information entry method provided in this embodiment of the invention can be applied to, for example... Figure 1 The application environment shown. Specifically, this information entry method can be applied in an information entry system, which includes, for example, […]. Figure 1 The diagram illustrates a client and server. The client and server communicate over a network to implement the steps of the information entry method. The client, also known as the user terminal, refers to the program that provides local services to the client, corresponding to the server. Clients can include, but are not limited to, various personal computers, laptops, smartphones, and tablets. The server can be a standalone server or a server cluster consisting of multiple servers.

[0014] Firstly, such as Figure 2 As shown, an information entry method is provided, which is applied to... Figure 1 Taking the server in the example, the following steps are included: S101: Obtain audio recording data, including dialogue data between the claims adjuster and the target injured party.

[0015] In this embodiment, the recorded data includes dialogue data between the claims adjuster and the target injured party. For example, it can be an audio file containing the dialogue data between the claims adjuster and the target injured party, acquired through a recording device (such as a voice recorder) or the audio acquisition module of the client. The recording data format can be common audio formats such as WAV, MP3, and AAC, and the sampling rate can be 8kHz, 16kHz, or 44.1kHz, etc. The claims adjuster refers to the insurance agent performing the visitation task. The target injured party refers to a natural person who is injured and receiving treatment in an insured event.

[0016] S102: Perform noise reduction processing on the recording data to obtain enhanced recording data.

[0017] In one possible approach, single-channel spectral subtraction noise reduction can be performed on the recording data to obtain enhanced recording data, thereby suppressing noise components and enhancing speech components, thus improving speech recognition clarity.

[0018] S103: Perform speaker segmentation and clustering, speech recognition, and speaker role labeling on the enhanced recording data in sequence to obtain role-labeled dialogue text. The role-labeled dialogue text is a global turn text sequence with role labels, including claims personnel and target injured persons.

[0019] In this embodiment, the server can sequentially perform speaker segmentation and clustering, speech recognition, and speaker role labeling on the enhanced recording data. Speaker segmentation and clustering refers to the technical processing of distinguishing and classifying the speech signals of different speakers in the enhanced recording data. In a visitation dialogue scenario, the enhanced recording data contains the speech signals of both the claims adjuster and the target injured party; the two speech signals alternate and are mixed and recorded in the same audio file. The server can extract the acoustic features of the enhanced recording data, analyze and cluster these features, and merge speech segments belonging to the same speaker to form at least two single-speaker speech streams, each containing the complete speech content of only one speaker.

[0020] Speech recognition is the technical process of converting speech signals from enhanced audio recordings into text data. Speech recognition can be implemented based on acoustic models and language models. Acoustic models map the acoustic features of speech signals to probability distributions of phonemes or words. Language models constrain and optimize the recognition results based on the linguistic rules of the context. The output data after speech recognition is a text sequence. This text sequence can contain multiple turns. A turn refers to a paragraph of text produced by a speaker in a single, continuous speech.

[0021] Speaker role labeling refers to classifying each turn in a text sequence by role, determining whether each turn is spoken by a claims adjuster or the target injured party. After speaker role labeling is completed, the resulting text is a role-labeled dialogue text. The role-labeled dialogue text is a global turn text sequence with role tags. The global turn text sequence refers to the complete dialogue text formed by arranging all turns in chronological order. Role tags are used to mark the identity category of the speaker corresponding to each turn.

[0022] S104: Input the character-annotated dialogue text into the preset information extraction model, and use the information extraction model to extract information and complete the address in the character-annotated dialogue text to obtain the claim visitation information of the target injured person output by the information extraction model. The claim visitation information includes personal information, work information, caregiver information and nursing information. Among them, personal information includes a first address and work information includes a second address. The first address and the second address are determined by the information extraction model after completing the incomplete addresses identified in the character-annotated dialogue text.

[0023] In one implementation, the server inputs the role-annotated dialogue text obtained in step S103 into a preset information extraction model. After receiving the role-annotated dialogue text as input, the information extraction model performs information extraction and address completion processing. During information extraction, the model uses the role tags marked on each turn in the role-annotated dialogue text to distinguish the information source. The model primarily extracts various types of information from the turn text tagged with "target injured person" and uses the contextual semantics provided by the turn text tagged with "claims personnel" to aid in understanding the target injured person's statements. The model performs semantic parsing on the input text, identifies the various information entities it contains, and categorizes and structures these entities according to preset information categories. During address completion, the model performs completion operations on incomplete addresses identified in the role-annotated dialogue text. The data output by the information extraction model is called claims visitation information.

[0024] The claims visitation information is a structured dataset containing multiple pre-defined information categories. These categories include personal information, employment information, dependent information, and care information. Personal information refers to information related to the injured party's identity and basic circumstances, and may include, but is not limited to, name, phone number, ID card number, and primary address. Employment information refers to information related to the injured party's occupation and income, and may include, but is not limited to, industry, employer name, secondary address, income details, and work start and end dates. Dependent information refers to information related to family members the injured party needs to support, and may include, but is not limited to, dependent's name, relationship, date of birth, duration of support, and number of dependents. Care information refers to information related to the care services the injured party receives, and may include, but is not limited to, type of care (e.g., home care, professional caregiver care, long-term care), income, employer, and fees.

[0025] In this context, the first address refers to the injured person's home address, and the second address refers to their work address. During visitation conversations, the injured person often only provides partial information about the address, such as the name of the residential complex, building, or street, omitting the complete administrative division information like province, city, and district. This type of address, lacking certain administrative division elements, is called an incomplete address. After identifying incomplete addresses in the character-annotated dialogue text, the information extraction model performs a completion operation. This completion operation supplements the missing administrative division information, transforming the incomplete address into a structured, complete address containing all necessary elements like province, city, district, and street. After the completion operation, the specific address content of the first address is determined by the completed home address, and the specific address content of the second address is determined by the completed work address.

[0026] As can be seen, the information extraction model in this embodiment extracts information from the target injured person's speech based on role tags, effectively avoiding the misinterpretation of the claims personnel's questions, restates, or confirmations as the target injured person's statements. The information extraction model intelligently completes incomplete addresses stated by the injured person, supplementing missing administrative division information. This eliminates the need for manual secondary queries and supplementation of the output results, directly meeting the claims system's requirement for address completeness and effectively reducing information distortion caused by incomplete address information.

[0027] It should be noted that the preset information extraction model is trained as follows: Historical visitation dialogue recordings are obtained; noise reduction is performed on these recordings to obtain enhanced recording samples; speaker segmentation and clustering, speech recognition, and speaker role labeling are performed on the enhanced recording samples to obtain role-labeled dialogue text samples; incomplete address text in the role-labeled dialogue text samples is identified and its type determined; online map services are used to complete the incomplete address text to obtain a complete address; the complete address is used to replace the corresponding incomplete address text in the role-labeled dialogue text samples to obtain role-labeled dialogue text samples containing complete addresses; structured claims information is extracted from the role-labeled dialogue text samples containing complete addresses as labeling information; the role-labeled dialogue text samples containing complete addresses are used as input samples, and the labeling information is used as labeling samples to construct a training sample set; the training sample set is then used to fine-tune the pre-trained large language model (such as Qwen3-235B-A22B) to obtain the information extraction model.

[0028] S105: Fill the corresponding fields of the claim visitation information into the corresponding claim case form.

[0029] In this embodiment, after the server obtains the claims visit information output by the information extraction model in step S104, it writes each information item in the claims visit information into the corresponding claims case form in the claims business system. The claims case form is an electronic data table or data page used in the claims business system to record detailed information of a single claims case. The claims case form defines multiple fields, each with a unique field identifier and a preset data type. The server stores a preset mapping table. This mapping table defines the correspondence between the names of each information item in the claims visit information and the field identifiers in the claims case form. Based on this mapping table, the server matches each information item in the claims visit information with the corresponding field in the claims case form. After matching, the server performs a fill operation. The fill operation refers to writing the specific data value of the information item into the storage location of the matched form field. After filling, the corresponding field in the claims case form is assigned the corresponding data content. After filling, claims personnel can view, verify, and confirm the information on the client interface.

[0030] As can be seen, this embodiment automatically writes structured claims visitation information into the claims case form, replacing the manual process of reviewing and entering information item by item, improving the efficiency of information entry, and effectively avoiding problems such as typing errors and field misalignment that may occur during manual entry, further reducing the problem of information input distortion.

[0031] In summary, the technical solution described in steps S101 to S105 above obtains enhanced recording data by denoising the audio data of the visitation dialogue. Then, speaker separation, speech recognition, and speaker role labeling are performed on the enhanced recording data to obtain role-labeled dialogue text. After inputting the role-labeled dialogue text into the information extraction model, the model extracts personal information, work information, caregiver information, and nursing information from the target injured person's turns of speech based on the role labels. It also completes any incomplete addresses identified, ultimately outputting structured claims visitation information and filling it into the corresponding fields of the corresponding claims form. Because the role-labeled dialogue text clearly identifies the speaker in each segment, the information extraction model can selectively extract information from the target injured person's statements, effectively avoiding the misinterpretation of claims personnel's questions or restatements as the target injured person's statements. Furthermore, the automatic extraction and filling of information replaces the manual input process after listening to the recording, further effectively avoiding information distortion caused by memory bias and improving the accuracy of information input.

[0032] In one embodiment, such as Figure 3 As shown, step S102, which involves noise reduction of the recording data to obtain enhanced recording data, includes the following steps: S121: Frame and window the recording data to obtain multiple audio frames; S122: Perform a short-time Fourier transform on each speech frame to obtain the spectrum of each speech frame; S123: In the silent section of the recording data, the noise estimator is used to estimate the spectrum of the silent section to obtain the noise spectrum; S124: Calculate the attenuation coefficient of each frequency band in each speech frame based on the spectrum and noise spectrum of each speech frame; S125: Perform spectral subtraction on the frequency bands with a signal-to-noise ratio less than a preset threshold in each speech frame, and retain the frequency bands with a signal-to-noise ratio greater than or equal to the preset threshold to obtain the spectral subtraction spectrum of each speech frame. S126: Perform inverse short-time Fourier transform on the spectrum of each speech frame after spectral subtraction to obtain enhanced recording data.

[0033] In this embodiment, framing refers to dividing continuous time-domain recording data into multiple short-time signal segments with a fixed frame length, which can be set to 20 milliseconds. Windowing multiplies the time-domain signal sample of each speech frame by a Hamming window function, the expression of which is: ,in For sample index, The total number of samples within a frame is used to reduce spectral leakage at frame boundaries. A short-time Fourier transform converts the windowed time-domain signal to the frequency domain, yielding the spectrum. A noise estimator statistically averages the spectral amplitudes of each frame during the silence segment to obtain the noise spectrum. The silence segment is determined by a speech activity detection algorithm based on short-time energy and zero-crossing rate. The attenuation coefficient is calculated based on the spectrum and noise spectrum of each frame; the calculation method can be... ,in For the first Attenuation coefficient of each frequency band This represents the noise spectrum amplitude of that frequency band. This represents the spectral amplitude of the speech frame in that frequency band. The signal-to-noise ratio (SNR) in each frequency band is lower than a preset threshold. A spectral subtraction operation is performed on the portion of the SNR below a preset threshold, multiplying the spectral amplitude by an attenuation coefficient to suppress noise. Frequency bands with an SNR at the preset threshold retain their amplitude unchanged. After spectral subtraction, an inverse short-time Fourier transform is performed on the subtracted spectrum to restore the signal from the frequency domain to the time domain. The frames are then stitched together using an overlap-addition method to reconstruct the complete enhanced recording data. Therefore, this embodiment effectively suppresses steady-state background noise components in the recording data and improves the SNR and clarity of the speech signal through frequency domain noise estimation and spectral subtraction.

[0034] In one embodiment, such as Figure 4 As shown, in step S103, which involves sequentially performing speaker segmentation and clustering, speech recognition, and speaker role labeling on the enhanced recording data to obtain the role-labeled dialogue text, the following steps are included: S131: Perform speaker segmentation and clustering on the enhanced recording data to obtain at least two single-person speech streams, each single-person speech stream corresponding to one speaker; S132: Perform speech recognition on each single-person speech stream to obtain the turn text sequence corresponding to each single-person speech stream. The turn text sequence contains multiple turns, and each turn is marked with a timestamp. S133: Arrange all the talk turns in chronological order according to the timestamp of each talk turn to obtain the global talk turn sequence; S134: Based on the topic succession relationship of each turn in the global turn sequence, determine whether each speaker is a claims adjuster or the target injured party; S135: Label each turn in the global turn sequence with its corresponding speaker role label to obtain the role-labeled dialogue text.

[0035] In one implementation, the server first performs speaker segmentation and clustering on the enhanced recording data. By extracting acoustic features from the enhanced recording data and detecting the moments when speaker identity changes in the audio, the enhanced recording data is segmented into multiple speech segments. Then, the speaker embedding vector of each speech segment is extracted, and a clustering algorithm is used to merge speech segments belonging to the same speaker, forming a single-speaker speech stream corresponding to that speaker. In visitation dialogue scenarios, there are usually two speakers, so after segmentation and clustering, two single-speaker speech streams are obtained, each containing the complete speech content of only one speaker. Next, the server performs speech recognition on each single-speaker speech stream. An acoustic model is used to map the speech signals in the single-speaker speech stream to a probability distribution of phonemes or words, and a language model is used to score and rank the candidate text sequences, outputting the turn text sequence corresponding to each single-speaker speech stream. Each turn contains the text content spoken by the speaker in a continuous speech, and is marked with the start and end times of that speech segment. Then, based on the timestamps of each turn, the server arranges all turns in the two turn-based text sequences in ascending order of their start times, resulting in a global turn sequence. This global turn sequence fully reconstructs the timeline and text content of the alternating speeches during the visitation dialogue. Finally, based on the topic progression of each turn in the global turn sequence, the server identifies each speaker as either an claims adjuster or the target injured party, and assigns a role label to each turn in the global turn sequence, resulting in role-labeled dialogue text. Therefore, this embodiment automatically converts the mixed dialogue audio into structured text with speaker identification, enabling the subsequent information extraction model to selectively extract the injured party's statements. This effectively avoids misinterpreting claims adjuster questions as injured party information, thereby reducing information distortion caused by confusion of information sources.

[0036] In one embodiment, such as Figure 5 As shown, in step S134, which involves determining whether each speaker is a claims adjuster or the target injured party based on the topic sequence of each turn in the global turn-taking sequence, the following steps are included: S1341: Calculate the time interval between the end time of each turn and the start time of the next turn in the global turn sequence; S1342: Determine adjacent turns with a time interval less than a preset time interval as a dialogue pair. The dialogue pair includes a preceding turn and a following turn. S1343: Calculate the semantic similarity between the text corresponding to the following turn and the text corresponding to the preceding turn; S1344: When the semantic similarity is greater than or equal to the preset semantic similarity, the subsequent turn is determined as the topic receiver of the preceding turn, and the preceding turn is determined as the topic initiator. S1345: When the semantic similarity is less than the preset semantic similarity, the subsequent turn will be determined as the topic initiator; S1346: Statistical analysis of the percentage of topic-initiating turns and the percentage of topic-responding turns for each speaker; S1347: Speakers who initiate conversations with a percentage greater than or equal to the first percentage are identified as claims personnel, and speakers who take over conversations with a percentage greater than or equal to the second percentage are identified as target injured persons.

[0037] In one possible implementation, the server first traverses the global turn sequence, calculating the time interval between the end time of each turn and the start time of the next turn. A preset time interval can be set to 2 seconds. The server identifies adjacent turns with a time interval less than 2 seconds as a dialogue pair, consisting of a preceding turn and a following turn. For each dialogue pair, the server uses a pre-trained language representation model to encode the preceding and following turn texts into semantic vectors, calculating the cosine similarity between the two semantic vectors as the semantic similarity value. A preset semantic similarity can be set to 0.6. When the cosine similarity is greater than or equal to 0.6, the server determines that the following turn semantically continues the topic of the preceding turn, marking the following turn as the topic receiver and the preceding turn as the topic initiator. When the cosine similarity is less than 0.6, the server determines that the following turn has started a new topic, marking the following turn as the topic initiator. After determining all dialogue pairs, the server calculates the percentage of each speaker's turns marked as topic initiators and the percentage of their turns marked as topic receivers. Both percentages can be set to 60%. The server identifies speakers whose topic initiator turns account for 60% or more as claims personnel and those whose topic receiver turns account for 60% or more as target injured parties. By quantifying time intervals and semantic similarity, role determination is based on objective and quantifiable data analysis, effectively avoiding the uncertainty of subjective judgment and significantly improving the accuracy and consistency of role determination.

[0038] For example, suppose that the enhanced recording data of a visitation conversation, after speaker segmentation and clustering, yields two single-speaker speech streams. After performing speech recognition on each of the two single-speaker speech streams, the following turn-taking text sequence is obtained: The first single-person speech stream corresponds to a text sequence containing four turns: Turn A1 starts at 00:02 and ends at 00:08, with the text "Where do you live now?"; Turn A2 starts at 00:16 and ends at 00:21, with the text "Where is your workplace?"; Turn A3 starts at 00:30 and ends at 00:35, with the text "How long have you been resting at home after your injury?"; Turn A4 starts at 00:44 and ends at 00:49, with the text "Do you have anyone in your family who needs support?" The second single-person speech stream corresponds to a text sequence containing four turns: Turn B1 starts at 00:09 and ends at 00:15, with the text "I live in X neighborhood of X district."; Turn B2 starts at 00:22 and ends at 00:29, with the text "I work at X company in Y district."; Turn B3 starts at 00:36 and ends at 00:43, with the text "I've been resting at home for almost three months."; Turn B4 starts at 00:50 and ends at 00:56, with the text "I have a son, born in June 2018." The server arranges all eight turns in ascending order of their start times based on their timestamps, resulting in the global turn sequence: Turn A1, Turn B1, Turn A2, Turn B2, Turn A3, Turn B3, Turn A4, Turn B4. This sequence reconstructs the temporal relationship of alternating "question and answer" interactions in the dialogue.

[0039] The server analyzes the topic succession relationship of the global turn sequence. The preset time interval is set to 2 seconds, and the preset semantic similarity is set to 0.6. The time interval between the end time of turn A1 (00:08) and the start time of turn B1 (00:09) is calculated to be 1 second, which is less than 2 seconds, thus turn A1 and turn B1 are identified as a dialogue pair. The semantic similarity between the text of turn B1 and the text of turn A1 is calculated, and the result is 0.85, which is greater than 0.6, thus turn B1 is determined to be the topic successor of turn A1, and turn A1 is the topic initiator. The time interval between the end time of turn B1 (00:15) and the start time of turn A2 (00:16) is calculated to be 1 second, which is less than 2 seconds, thus turn B1 and turn A2 are identified as a dialogue pair. The semantic similarity between the text of turn A2 and the text of turn B1 is calculated, and the result is 0.28, which is less than 0.6, thus turn A2 is determined to be the topic initiator. In this way, adjacent turns in the global turn sequence are analyzed one by one to complete the determination of all dialogue pairs.

[0040] Statistical analysis revealed that the first single-person voice stream had 4 topic initiation turns and 0 topic follow-up turns, resulting in a 100% topic initiation rate, exceeding the first stream's 60% rate. Therefore, this speaker was identified as a claims adjuster. The second single-person voice stream also had 0 topic initiation turns and 4 topic follow-up turns, resulting in a 100% topic follow-up rate, exceeding the second stream's 60% rate. Therefore, this speaker was identified as the target injured party. The server labeled each turn in the global turn sequence with a corresponding role tag: turns A1 to A4 were labeled as claims adjusters, and turns B1 to B4 were labeled as target injured parties, resulting in the role-labeled dialogue text.

[0041] It should be noted that the above is merely an example and does not constitute a real processing procedure, nor does it limit the embodiments of the present invention.

[0042] In one embodiment, such as Figure 6 As shown, in step S104, where the first and second addresses are determined by the information extraction model after completing the incomplete addresses identified in the character-annotated dialogue text, the steps include: S141A: The information extraction model identifies incomplete addresses from the turn text labeled as the target injured person in the role-annotated dialogue text; S142A: The information extraction model determines the address type corresponding to the incomplete address based on the contextual semantics of the turn where the incomplete address is located and the question semantics of the turn labeled as a claims person regarding the incomplete address. The address type includes home address or work address. S143A: The information extraction model uses an incomplete address as the query term, calls an online map service to perform a query, obtains a list of candidate addresses, and selects a candidate address that matches the address type from the list as the completed address; when the address type is a home address, the information extraction model uses the completed address as the first address; when the address type is a work address, the information extraction model uses the completed address as the second address.

[0043] In one implementation, the named entity recognition component within the information extraction model performs word-by-word sequence annotation on the turn text labeled as the target injured person in the role-annotated dialogue text, identifying incomplete address text lacking complete administrative division information such as province, city, and district. The intent classification component obtains the preceding turn question text labeled as the claims person, encodes it into a semantic vector, and classifies it within a predefined intent category set. When the classification result is "inquiring about home address," the address type corresponding to the incomplete address text is determined as a home address; when the classification result is "inquiring about work address," it is determined as a work address. The information extraction model constructs a Hypertext Transfer Protocol request, using the incomplete address text as a query term parameter, and calls the application programming interface of the online map service. After the online map service returns a list of candidate addresses, the information extraction model selects the matching candidate address as the completed address based on the determined address type, and uses the completed home address as the first address field of the personal information, or the completed work address as the second address field of the work information.

[0044] For example, the character-annotated dialogue text of a visitation conversation contains the following turns: In the first round of conversation, the character is tagged as an claims adjuster, and the text is "Where do you currently live?" Turn 2, the character is tagged as the injured person, and the text is "I live in the X neighborhood of the X district". In turn 3, the character is tagged as an claims adjuster, and the text is "Where does your workplace lie?" Turn 4, the character is tagged as the target injured person, and the text content is "I work at company X in district Y". The information extraction model receives the above-mentioned character-annotated dialogue text as input and performs address completion processing.

[0045] In S141A, the named entity recognition component of the information extraction model performs sequence labeling on the second statement text "I live in the x community of x district" labeled as the target injured person, marking "x community of x district" as an address entity. This address entity lacks complete administrative division information such as province and city, and is marked as incomplete address text. Sequence labeling is performed on the fourth statement text "I work at x company in y district" labeled as the target injured person, marking "x company in y district" as an address entity, which is also marked as incomplete address text due to the lack of complete administrative division information.

[0046] In S142A, for the incomplete address text "x area x neighborhood", the intent classification component obtains the question text from the preceding talk round 1, "Where do you live now?", and outputs the category "Inquiry about home address", thus determining that the address type corresponding to the incomplete address text is a home address. For the incomplete address text "y area x company", the intent classification component obtains the question text from the preceding talk round 3, "Where is your workplace?", and outputs the category "Inquiry about work address", thus determining that the address type corresponding to the incomplete address text is a work address.

[0047] In S143A, the information extraction model uses "x community" as the query term to call the online map service, returning multiple candidate addresses, including "x province x city x district x community" and the same-named x community located in other provinces and cities. The information extraction model has already extracted the work address entity "x province x city y district x company" from the text of turn 4, using it as a spatial reference anchor point. It calculates the geographical distance between each candidate address and this anchor point, selecting the candidate address with the smallest distance, "x province x city x district x community", as the completed address. Since the address type is a home address, the completed address is used as the first address. For the incomplete address text "y district x company", the information extraction model uses "x company" as the query term to call the online map service, returning the candidate address "x province x city y district x company", which matches the address type "work address". This candidate address is selected as the completed address and used as the second address.

[0048] As can be seen, through the above processing, the named entity recognition component, intent classification component, and online map calling component within the information extraction model work together to automatically complete incomplete, colloquial addresses in the dialogue into complete, structured addresses, eliminating the need for manual querying and filling in, thereby reducing information distortion caused by address completion errors.

[0049] In one embodiment, such as Figure 7 As shown, in step S104, the job information also includes monthly income information; wherein, the monthly income information is determined by the information extraction model in the following way, including the following steps: S141B: Identify income description text containing fuzzy quantifiers and income value ranges from the turn text labeled as the target injured person in the character-annotated dialogue text; S142B: Extract the upper and lower bounds of the income value range from the income description text, calculate the median of the upper and lower bounds, and use the median as the monthly income information.

[0050] In one implementation, a fuzzy semantic recognition component within the information extraction model analyzes the dialogue text labeled as belonging to the target injured party in role-annotated dialogue text. This component uses a combination of regular expression pattern matching and keyword matching to identify income description text that simultaneously contains fuzzy quantifiers and income value ranges. Fuzzy quantifiers refer to words expressing uncertainty, such as "approximately," "around," or "more or less." Income value ranges refer to a numerical range formed by two consecutive values ​​in spoken language. After identifying the income description text, the value range extraction component within the information extraction model extracts the upper and lower bounds of the income value range in the text. The upper bound is the larger value in the range, and the lower bound is the smaller value. After the value range is extracted, the information extraction model calculates the median of the upper and lower bounds by adding them together and dividing by two. The information extraction model uses this calculated median as the monthly income information for the work information in the claims visitation information.

[0051] For example, the character-annotated dialogue text of a visitation conversation contains the following turns: In turn 1, the character is tagged as an claims adjuster, and the text is "What is your approximate monthly income?" In the second round of conversation, the character is tagged as the target wounded, and the text reads, "Seven or eight thousand a month, it's not fixed." The information extraction model receives the aforementioned character-annotated dialogue text as input and performs monthly income information determination processing. In S141B, the fuzzy semantic recognition component of the information extraction model analyzes the text "seven or eight thousand a month, not very fixed" in turn 2, labeled as the target injured person. Through regular expression pattern matching, it identifies the text fragment "seven or eight thousand" containing two consecutive numerical values; through keyword matching, it identifies the fuzzy quantifiers "ba" and "not very fixed" in the text. The fuzzy semantic recognition component determines that the text fragment contains both fuzzy quantifiers and an income value range, and marks it as income description text. In S142B, the numerical range extraction component of the information extraction model performs numerical parsing on the identified income description text "seven or eight thousand", extracting the lower bound of the income value range as 7000, the upper bound as 8000, and calculating the median as 7500. The information extraction model uses the median of 7500 as the monthly income information, and uses it as the monthly income information for the work information in the claims visitation information.

[0052] As can be seen, through the above processing, the fuzzy semantic recognition component and the numerical range extraction component inside the information extraction model work together to automatically quantify the colloquial and vague income description in the dialogue into precise numerical information. This avoids the inconsistency and error caused by subjective estimation by humans when they hear vague expressions, and ensures the uniformity of the monthly income information quantification standard and the objectivity and accuracy of the input results.

[0053] In one embodiment, such as Figure 8 As shown, after step S105, the nursing information includes the standard fee for caregivers; that is, after filling the corresponding field of the claim visitation information into the corresponding claim form, the following steps are also included: S106: Obtain the administrative region to which the first address or the second address belongs; S107: Obtain a reference range for caregiver fees within the administrative region based on historical claims data statistics; S108: Compare the standard caregiver fee with the reference range for caregiver fees; S109: When the caregiver fee standard value exceeds the upper limit of the caregiver fee standard reference range, a risk mark is generated, and the display color of the caregiver fee standard value in the claim form is marked according to the preset color corresponding to the risk mark.

[0054] In one possible implementation, the server extracts the text of the completed first address from the claims visitation information. It then breaks down the address text level by level using preset administrative division resolution rules, extracting the district / county-level administrative division name as the administrative region. Using this administrative region as a query condition, the server retrieves historical caregiver fee data records for that region from the historical claims database. It performs statistical analysis on the extracted numerical set to calculate the normal fluctuation range of caregiver fees within that administrative region. The upper limit of the reference interval is set to the 95th percentile of the numerical set. The server compares the current case's caregiver fee standard value with the upper limit of the reference interval. When the caregiver fee standard value exceeds the upper limit, the server generates a risk marker data structure containing risk type and risk level information. When generating the rendering data for the claims form, the server applies a color style attribute to the display area of ​​the caregiver fee standard value based on the preset red color corresponding to the risk marker. This value is then highlighted in red during client rendering.

[0055] For example, suppose the first address in a claim visitation message, after completion, is "Province x, City x, District x, Community x", and the standard caregiver fee is x yuan per day. The server extracts the administrative region as "District x" from the first address. It queries the historical claims database for historical caregiver fee records for "District x", and through statistical analysis, obtains a reference range of m yuan to n yuan per day, with an upper limit of n yuan per day. The current standard caregiver fee of x yuan per day exceeds the upper limit of n yuan per day, so the server generates a risk marker and marks the value in red. When claims personnel view the form on the client, they can quickly identify the abnormal risk of this value through the red highlighting and perform manual verification and correction.

[0056] It can be seen that, through the above processing in this embodiment, the risk identification logic is embedded into the information entry processing flow, and the statistical characteristics of historical data are used to automatically verify key values and perform abnormal warning. When the value of the caregiver's charging standard exceeds the normal range of the region, the system automatically generates a risk mark and intuitively prompts the claim adjuster in a color marking manner, which realizes automatic identification and visual warning of abnormal data, and assists in improving the efficiency and coverage of claim risk control.

[0057] In an embodiment, that is, in step S104, the dependent information includes the dependency duration; wherein the dependency duration is determined by the information extraction model in the following manner: S141D: The information extraction model identifies the date of birth of the dependent from the turn text labeled as the target injured person in the role-labeled dialogue text. S142D: The information extraction model calculates the remaining dependency duration based on the current date and the preset legal dependency duration rule, and takes the remaining dependency duration as the dependency duration.

[0058] In an implementable manner, the date recognition component inside the information extraction model analyzes the turn text labeled as the target injured person in the role-labeled dialogue text. The date recognition component adopts a date pattern matching algorithm to scan the turn text of the target injured person word by word, so as to identify text fragments conforming to the date format. Recognizable date formats include common date expressions such as "YYYY年MM月", "YYYY-MM-DD", "YYYY.MM.DD" and the like. The date recognition component parses the matched date text fragment into structured date data including a year field, a month field and a date field, and marks the date data as the date of birth of the dependent. The information extraction model obtains the current date through the system clock. A legal dependency duration rule is preset inside the information extraction model, and the preset value is that the dependent reaches the age of 18. The information extraction model calculates the current age of the dependent based on the current date and the date of birth, and the calculation method is: subtracting the birth year from the current year to obtain a preliminary age difference, then comparing the relationship between the current month and day and the birth month and day; if the current date has not reached the birth month and day, the preliminary age difference is subtracted by one, indicating that the full age has not been reached. The calculation method of the remaining dependency duration is: subtracting the current age of the dependent from the legal dependency duration of 18 years old. The information extraction model assigns the calculated remaining dependency duration value to the dependency duration information field of the dependent information in the claim visit information.

[0059] Exemplarily, the role-labeled dialogue text of a certain visit dialogue includes the following turns: Turn 1, with the role label of claim adjuster, the text content is "Do you have any children in need of dependency at home?" Turn 2, with the role label of target injured person, the text content is "I have a son, who was born in June 2018." The information extraction model receives the aforementioned character-annotated dialogue text as input and performs processing to determine the dependency period information. In S141D, the date recognition component of the information extraction model performs date pattern matching on the text "has a son, born in June 2018" in turn 2 labeled as the target injured person, identifying the text fragment "June 2018" in the format "YYYY year MM month". The date recognition component parses this text fragment into structured date data, with the year field being 2018 and the month field being 6, and marks it as the dependent's birth date. In S142D, the information extraction model obtains the current date as May 2026 from the system clock. According to the preset legal dependency period rule, the dependency period is 18 years old. The information extraction model calculates the dependent's current age: subtracting the birth year 2018 from the current year 2026, the initial age difference is 8 years; since the current month May has not yet reached the birth month June, the initial age difference is reduced by one, and the dependent's current age is 7 years old. The information extraction model subtracts 7 years from 18 years old to calculate the remaining support period as 11 years. The information extraction model then uses this remaining support period of 11 years as the support period information field in the supporter information section of the claims visitation information.

[0060] As can be seen, through the above processing, the date recognition component inside the information extraction model automatically parses the colloquial birth date description in the dialogue into structured date data, and automatically calculates the remaining years of support based on the legal support period rule, without the need for manual calendar lookup and calculation, thus improving the accuracy and processing efficiency of the support period information entry.

[0061] It should be noted that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0062] The technical solution of this invention will be illustrated below using a specific financial insurance claims visitation application scenario.

[0063] An insurance company claims adjuster received a visitation assignment for a car insurance personal injury claim and was tasked with visiting a traffic accident victim at a hospital. The adjuster arrived at the hospital carrying a smartphone with the insurance company's claims processing application installed. In the patient's ward, the adjuster spoke face-to-face with the victim and began recording the conversation using the application's recording function. The conversation is as follows: Claims adjuster: "Hello, I'm a claims adjuster from the insurance company. I'm here today to understand your injuries and some basic information. What is your contact number?" The injured person said, "My phone number is 13xxxx456xx." Claims adjuster: "Okay, what is your ID number?" The injured person's ID number is: "4403xx19850615xxxx". Claims adjuster: "Where do you live now?" The injured person said, "I live in the X neighborhood of the X district." Claims adjuster: "Where does your workplace lie?" The injured party said, "I work at Company X in District Y." Claims adjuster: "How long have you been working there?" The injured person said, "It's been more than three years." Claims adjuster: "What is your approximate monthly income?" The injured person said, "It's about seven or eight thousand a month, but it's not very fixed." Claims adjuster: "Have you been resting at home since you were injured?" The injured person said, "Yes, I've been resting at home for almost three months." Claims adjuster: "Do you have any children or elderly relatives in your family who need support?" The injured party said, "I have a son who was born in June 2018." Claims adjuster: "Who took care of you while you were injured?" The injured person said, "I hired a caregiver who charges 300 yuan a day." Claims officer: "Okay, I have recorded all the above information. Please check it." After the claims personnel complete the recording, they upload the recording data to the insurance company's server via the application. Upon receiving the recording data, the server executes the information entry processing procedure.

[0064] The server first performs noise reduction on the recorded data. The ward environment contains background noise such as air conditioning operation and medical equipment prompts. The server segments the recorded data into 20-millisecond speech frames, applies a Hamming window, and performs a short-time Fourier transform to obtain the spectrum of each frame. During silent segments before recording begins and between conversations, a noise estimator is used to extract the noise spectrum. The attenuation coefficient for each frequency band in each frame is calculated, and spectral subtraction is performed on frequency bands with a signal-to-noise ratio (SNR) below a preset threshold, retaining frequency bands with an SNR at the preset threshold. An inverse short-time Fourier transform is then performed on the subtracted spectrum to reconstruct the enhanced recorded data. After noise reduction, background noise is significantly suppressed, and the speech is clearer.

[0065] Next, the server performs speaker segmentation and clustering, speech recognition, and speaker role labeling on the enhanced recording data. Based on differences in acoustic features, the server separates the recording into two single-speaker speech streams, corresponding to speaker A and speaker B respectively. Speech recognition is performed on each single-speaker speech stream to obtain the turn text sequence for each stream. All turns are arranged by timestamp to obtain the global turn sequence. The time interval between adjacent turns is calculated, and adjacent turns with an interval of less than 2 seconds are identified as dialogue pairs. The semantic similarity between the subsequent turn and the preceding turn in each dialogue pair is calculated. Statistically, speaker A initiated 90% of the turns, exceeding the first speaker's 60%, thus speaker A is identified as the claims adjuster; speaker B received 88% of the turns, exceeding the second speaker's 60%, thus speaker B is identified as the injured party. Each turn is labeled with a corresponding role tag, resulting in role-labeled dialogue text.

[0066] Then, the server inputs the character-annotated dialogue text into the information extraction model. The information extraction model performs semantic parsing and information extraction on the character-annotated dialogue text to generate structured claims visit information. When extracting address information, the information extraction model identifies incomplete address text from the dialogue of the person labeled as injured, determines the address type as home address or work address based on the semantics of the claims personnel's question, and calls the online map service using the incomplete address text as the query term. When the map service returns multiple candidate addresses, the extracted work address entity is used as a spatial reference anchor point, and the nearest candidate address is selected as the completed address by calculating the geographical distance. When extracting monthly income information, the income description text "seven or eight thousand a month" is identified, the lower bound value of 7000 and the upper bound value of 8000 are extracted, and the median of 7500 is calculated as the monthly income information. When extracting lost work time information, the colloquial description text "I've been resting at home for almost three months" is identified, and the number of lost work days is obtained by parsing the audio recording timestamp. When extracting information on caregivers, the system identifies the dependent's birth date as "June 2018" and calculates the remaining years of support based on the current date and the legal support period rule of 18 years of age. When extracting information on caregivers, the system identifies the caregiver's fee as 300 yuan per day.

[0067] The information extraction model outputs complete structured visitation information for claims, including personal information, work information, dependent information, and care information. The server populates the corresponding fields in the claim form with this visitation information. Simultaneously, the server performs a risk check on the caregiver fee standard. The historical reference range for caregiver fees in the administrative region of the first address is 150 to 280 yuan per day. The current value of 300 yuan per day exceeds the upper limit, generating a risk marker and highlighting it in red. When the claims personnel view the form on the application and find the highlighted value, after confirming with the injured party, they manually correct the value to 280 yuan per day.

[0068] In summary, through the above process, the recorded visitation conversations are automatically processed into structured claim visitation information and accurately filled into the form within minutes. Claims personnel only need to verify and make minor corrections to complete the data entry. Throughout the process, the system automatically distinguishes between the statements of the claims personnel and the injured party, accurately locates the information provided by the injured party, automatically completes incomplete addresses, effectively avoiding manual inquiries, automatically quantifies colloquial descriptions of income and lost work time, effectively avoiding errors in manual estimation, and provides intelligent risk warnings for abnormal caregiver charges. Data entry efficiency is significantly improved, the accuracy and consistency of the entered information are effectively guaranteed, and the problem of information distortion caused by manual entry is reduced.

[0069] Secondly, an information input device is provided, which corresponds one-to-one with the information input methods in the above embodiments. For example... Figure 9 As shown, the information input device includes an acquisition module 201, a noise reduction module 202, a differentiation module 203, an extraction module 204, and a filling module 205. Detailed descriptions of each functional module are as follows: The acquisition module 201 is used to acquire audio recording data, which includes dialogue data between the claims adjuster and the target injured party; Noise reduction module 202 is used to perform noise reduction processing on the recording data to obtain enhanced recording data; The differentiation module 203 is used to sequentially perform speaker segmentation and clustering, speech recognition and speaker role labeling on the enhanced recording data to obtain role-labeled dialogue text. The role-labeled dialogue text is a global turn text sequence with role labels, and the role labels include claims personnel and target injured persons. The extraction module 204 is used to input the character-annotated dialogue text into a preset information extraction model, and use the information extraction model to perform information extraction and address completion processing on the character-annotated dialogue text to obtain the claims visitation information of the target injured person output by the information extraction model. The claims visitation information includes personal information, work information, caregiver information, and nursing information; wherein, the personal information includes a first address, the work information includes a second address, and the first address and the second address are determined by the information extraction model after completing the incomplete addresses identified in the character-annotated dialogue text. The fill module 205 is used to fill the claim visitation information into the corresponding fields of the corresponding claim case form.

[0070] In one embodiment, the noise reduction module 202 is further configured to: perform frame-by-frame windowing on the recording data to obtain multiple speech frames; perform a short-time Fourier transform on each speech frame to obtain the spectrum of each speech frame; estimate the spectrum of the silent segment in the recording data using a noise estimator to obtain a noise spectrum; calculate the attenuation coefficient of each frequency band in each speech frame based on the spectrum of each speech frame and the noise spectrum; perform a spectral subtraction operation on the frequency bands in each speech frame where the signal-to-noise ratio is less than a preset threshold, and retain the frequency bands where the signal-to-noise ratio is greater than or equal to the preset threshold to obtain the spectral subtracted spectrum of each speech frame; and perform an inverse short-time Fourier transform on the spectral subtracted spectrum of each speech frame to obtain the enhanced recording data.

[0071] In one embodiment, the differentiation module 203 is further configured to: perform speaker segmentation and clustering on the enhanced recording data to obtain at least two single-person speech streams, each single-person speech stream corresponding to one speaker; perform speech recognition on each single-person speech stream to obtain a turn text sequence corresponding to each single-person speech stream, the turn text sequence containing multiple turns, each turn being timestamped; arrange all turns in chronological order according to the timestamps of each turn to obtain a global turn sequence; determine whether each speaker is an claims adjuster or a target injured party according to the topic sequence of each turn in the global turn sequence; and label each turn in the global turn sequence with its corresponding speaker role tag to obtain the role-labeled dialogue text.

[0072] In one embodiment, the differentiation module 203 is further configured to: calculate the time interval between the end time of each turn and the start time of the next turn in the global turn sequence; determine adjacent turns with a time interval less than a preset time interval as a dialogue pair, the dialogue pair including a preceding turn and a following turn; calculate the semantic similarity between the text corresponding to the following turn and the text corresponding to the preceding turn; when the semantic similarity is greater than or equal to a preset semantic similarity, determine the following turn as the topic receiver of the preceding turn and the preceding turn as the topic initiator; when the semantic similarity is less than the preset semantic similarity, determine the following turn as the topic initiator; count the percentage of topic initiating turns and the percentage of topic receiving turns for each speaker; determine speakers with a percentage of topic initiating turns greater than or equal to a first percentage as claims personnel, and determine speakers with a percentage of topic receiving turns greater than or equal to a second percentage as target injured persons.

[0073] In one embodiment, the extraction module 204 is further configured to perform the following operations using the information extraction model: identify incomplete addresses from the dialogue text labeled as the target injured person in the role-annotated dialogue text; determine the address type corresponding to the incomplete address based on the contextual semantics of the dialogue text containing the incomplete address and the question semantics of the dialogue text labeled as the claims personnel regarding the incomplete address, wherein the address type includes a home address or a work address; query an online map service using the incomplete address as a query term to obtain a candidate address list, and select a candidate address matching the address type from the candidate address list as the completed address; wherein, when the address type is a home address, the completed address is used as the first address; when the address type is a work address, the completed address is used as the second address.

[0074] In one embodiment, the extraction module 204 is further configured to perform the following operations through the information extraction model: identify income description text containing fuzzy quantifiers and income value ranges from the turn text labeled as the target injured person in the character-annotated dialogue text; extract the upper and lower bound values ​​of the income value range in the income description text, calculate the median of the upper and lower bound values, and use the median as the monthly income information.

[0075] In one embodiment, the information entry device further includes a risk marking module 206, configured to: obtain the administrative region to which the first address or the second address belongs; obtain a reference range of caregiver fee standards based on historical claims data within the administrative region; compare the caregiver fee standard value in the nursing information with the caregiver fee standard reference range; when the caregiver fee standard value exceeds the upper limit of the caregiver fee standard reference range, generate a risk mark, and mark the display color of the caregiver fee standard value in the claims form according to the preset color corresponding to the risk mark.

[0076] Specific limitations regarding the information entry device can be found in the limitations on the information entry method described above, and will not be repeated here. Each module in the aforementioned information entry device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in the computer device in hardware form, or stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0077] Thirdly, a computer device is provided, which can be a server, and its internal structure diagram can be as follows: Figure 10As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When the computer program is executed by the processor, it implements the steps of an information entry method.

[0078] In one embodiment, a computer device is provided, 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 perform the following steps: Acquire audio recordings, including conversations between claims adjusters and the target injured party; The audio data is subjected to noise reduction processing to obtain enhanced audio data; The enhanced recording data is sequentially subjected to speaker segmentation and clustering, speech recognition, and speaker role labeling to obtain role-labeled dialogue text. The role-labeled dialogue text is a global turn-taking text sequence with role labels, and the role labels include claims personnel and target injured persons. The character-annotated dialogue text is input into a preset information extraction model. The information extraction model is used to extract information and complete addresses from the character-annotated dialogue text to obtain the claims visitation information of the target injured person output by the information extraction model. The claims visitation information includes personal information, work information, caregiver information, and nursing information. Among them, the personal information includes a first address, and the work information includes a second address. The first address and the second address are determined by the information extraction model after completing the incomplete addresses identified in the character-annotated dialogue text. Fill the corresponding fields of the claim visitation information into the corresponding claim case form.

[0079] Fourthly, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, performs the following steps: Acquire audio recordings, including conversations between claims adjusters and the target injured party; The audio data is subjected to noise reduction processing to obtain enhanced audio data; The enhanced recording data is sequentially subjected to speaker segmentation and clustering, speech recognition, and speaker role labeling to obtain role-labeled dialogue text. The role-labeled dialogue text is a global turn-taking text sequence with role labels, and the role labels include claims personnel and target injured persons. The character-annotated dialogue text is input into a preset information extraction model. The information extraction model is used to extract information and complete addresses from the character-annotated dialogue text to obtain the claims visitation information of the target injured person output by the information extraction model. The claims visitation information includes personal information, work information, caregiver information, and nursing information. Among them, the personal information includes a first address, and the work information includes a second address. The first address and the second address are determined by the information extraction model after completing the incomplete addresses identified in the character-annotated dialogue text. Fill the corresponding fields of the claim visitation information into the corresponding claim case form.

[0080] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0081] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0082] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0083] It should be noted that if any AI models, software tools, or components not belonging to this company appear in the embodiments of this invention, they are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this invention is authorized (with the knowledge and consent) by the relevant parties or fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.

[0084] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. An information entry method, characterized in that, include: Acquire audio recordings, including conversations between claims adjusters and the target injured party; The audio data is subjected to noise reduction processing to obtain enhanced audio data; The enhanced recording data is sequentially subjected to speaker segmentation and clustering, speech recognition, and speaker role labeling to obtain role-labeled dialogue text. The role-labeled dialogue text is a global turn-taking text sequence with role labels, and the role labels include claims personnel and target injured persons. The character-annotated dialogue text is input into a preset information extraction model. The information extraction model is used to extract information and complete addresses from the character-annotated dialogue text to obtain the claims visitation information of the target injured person output by the information extraction model. The claims visitation information includes personal information, work information, caregiver information, and nursing information. Among them, the personal information includes a first address, and the work information includes a second address. The first address and the second address are determined by the information extraction model after completing the incomplete addresses identified in the character-annotated dialogue text. Fill the corresponding fields of the claim visitation information into the corresponding claim case form.

2. The method according to claim 1, characterized in that, The noise reduction process on the recorded data to obtain enhanced recorded data includes: The recorded data is segmented and windowed to obtain multiple audio frames; Perform a short-time Fourier transform on each of the speech frames to obtain the spectrum of each speech frame; In the silent segment of the recorded data, the frequency spectrum of the silent segment is estimated using a noise estimator to obtain the noise spectrum; Based on the spectrum of each speech frame and the noise spectrum, calculate the attenuation coefficient of each frequency band in each speech frame; A spectral subtraction operation is performed on the frequency bands in each of the speech frames where the signal-to-noise ratio is less than a preset threshold, and the frequency bands with a signal-to-noise ratio greater than or equal to the preset threshold are retained to obtain the spectral subtraction spectrum of each of the speech frames; The enhanced recording data is obtained by performing an inverse short-time Fourier transform on the spectral subtraction spectrum of each of the aforementioned speech frames.

3. The method according to claim 1, characterized in that, The enhanced recording data is sequentially subjected to speaker segmentation and clustering, speech recognition, and speaker role labeling to obtain role-labeled dialogue text, including: The enhanced recording data is segmented and clustered by speaker to obtain at least two single-person speech streams, each of which corresponds to one speaker. Speech recognition is performed on each of the single-person speech streams to obtain the turn text sequence corresponding to each of the single-person speech streams. The turn text sequence contains multiple turns, and each turn is marked with a timestamp. Based on the timestamps of each talk turn, all talk turns are arranged in chronological order to obtain the global talk turn sequence; Based on the topic sequence of each turn in the global turn-taking sequence, each speaker is identified as either a claims adjuster or the target injured party. Each turn in the global turn sequence is labeled with its corresponding speaker role tag to obtain the role-labeled dialogue text.

4. The method according to claim 3, characterized in that, The step of determining whether each speaker is a claims adjuster or a target injured party based on the topic sequence of each turn in the global turn-taking sequence includes: Calculate the time interval between the end time of each turn and the start time of the next turn in the global turn sequence; Adjacent turns with a time interval less than a preset time interval are defined as a dialogue pair, and the dialogue pair includes a preceding turn and a following turn; Calculate the semantic similarity between the text corresponding to the subsequent turn and the text corresponding to the preceding turn; When the semantic similarity is greater than or equal to the preset semantic similarity, the subsequent turn is determined as the topic receiver of the preceding turn, and the preceding turn is determined as the topic initiator. When the semantic similarity is less than the preset semantic similarity, the subsequent turn is identified as the topic initiator; The percentage of topic-initiating turns and the percentage of topic-responding turns for each speaker were statistically analyzed. Speakers who initiate conversations at a rate greater than or equal to the first percentage are identified as claims personnel, and speakers who take over conversations at a rate greater than or equal to the second percentage are identified as target injured persons.

5. The method according to claim 1, characterized in that, The first address and the second address are determined by the information extraction model after completing the incomplete addresses identified in the character-annotated dialogue text, including: The information extraction model identifies incomplete addresses from the dialogue text labeled as the target injured person in the character-annotated dialogue text; The information extraction model determines the address type corresponding to the incomplete address based on the contextual semantics of the turn where the incomplete address is located and the question semantics of the turn labeled as a claims person regarding the incomplete address. The address type includes home address or work address. The information extraction model uses the incomplete address as a query term, calls an online map service to perform a query, obtains a candidate address list, and selects a candidate address that matches the address type from the candidate address list as the completed address; wherein, when the address type is a home address, the information extraction model uses the completed address as the first address; when the address type is a work address, the information extraction model uses the completed address as the second address.

6. The method according to claim 1, characterized in that, The job information also includes monthly income information; the monthly income information is determined by the information extraction model in the following way: From the dialogue text labeled as the target injured person in the character-annotated dialogue text, income description text containing fuzzy quantifiers and income value ranges was identified; Extract the upper and lower bound values ​​of the income value range from the income description text, calculate the median of the upper and lower bound values, and use the median as the monthly income information.

7. The method according to claim 1, characterized in that, The nursing information includes the standard fee for caregivers; after filling the visitation information into the corresponding field of the corresponding claim form, the method further includes: Obtain the administrative region to which the first address or the second address belongs; Obtain a reference range for caregiver fees within the administrative region based on historical claims data. Compare the stated caregiver fee standard with the stated caregiver fee standard reference range; When the caregiver fee standard value exceeds the upper limit of the caregiver fee standard reference range, a risk mark is generated, and the display color of the caregiver fee standard value in the claim form is marked according to the preset color corresponding to the risk mark.

8. An information input device, characterized in that, include: The acquisition module is used to acquire audio data, which includes dialogue data between the claims adjuster and the target injured party; A noise reduction module is used to perform noise reduction processing on the recording data to obtain enhanced recording data; The differentiation module is used to sequentially perform speaker segmentation and clustering, speech recognition, and speaker role labeling on the enhanced recording data to obtain role-labeled dialogue text. The role-labeled dialogue text is a global turn-taking text sequence with role labels, and the role labels include claims personnel and target injured persons. An extraction module is used to input the character-annotated dialogue text into a preset information extraction model, and use the information extraction model to perform information extraction and address completion processing on the character-annotated dialogue text to obtain the claims visitation information of the target injured person output by the information extraction model. The claims visitation information includes personal information, work information, caregiver information, and nursing information; wherein, the personal information includes a first address, the work information includes a second address, and the first address and the second address are determined by the information extraction model after completing the incomplete addresses identified in the character-annotated dialogue text. The fill module is used to fill the claim visitation information into the corresponding fields of the corresponding claim case form.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the information entry method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the information entry method as described in any one of claims 1 to 7.

Images