Methods, devices, computer equipment, and storage media for identifying target content in live streaming rooms

By performing multi-dimensional verification on live audio and video data, generating structured corpus and identifying commitment information, the problem of accuracy in commitment information identification in live marketing is solved, the false detection rate is reduced, and the authenticity assessment of commitment information is improved.

CN123070132APending Publication Date: 2026-11-10JINGUAZI TECH DEV CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611586036.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-14
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

In existing live-streaming marketing, the identification of the promises made by the anchors lacks accuracy. Traditional NLP methods suffer from insufficient recognition capabilities, fragmented output formats of large models, frequent model illusions, and high recall rates of false recognition, leading to difficulties in subsequent performance management and customer complaint handling.

Method used

By acquiring live audio and video data, identifying and fusing them to generate role tags, processing text data by slicing and associating it with timestamps and contextual information, and using multi-dimensional verification methods including original text anchoring, contextual consistency and business constraint verification, the system identifies commitment information by combining preset prompt word templates and performs output control processing.

Benefits of technology

It enables accurate identification and evaluation of commitment information during live streaming, reduces false detection rate, improves the evaluation of the authenticity and validity of commitment information, and reduces the recall of invalid commitments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123070132A_ABST
    Figure CN123070132A_ABST
Patent Text Reader

Abstract

This application provides a method, apparatus, computer device, and storage medium for identifying target content in a live stream, applied in the field of artificial intelligence. The method includes: after generating role tags, converting live stream audio data into live stream text data, slicing the live stream text data, and associating the sliced ​​text data segments with role tags, timestamps, and conversation context information to generate structured corpus; performing model recognition on the structured corpus based on various promise prompts from a preset prompt word template to determine promise information; performing multi-dimensional verification on the promise information to determine multi-dimensional verification scores, including original text anchoring verification, contextual consistency verification, and business constraint verification; and performing output control processing on the promise information based on the multi-dimensional verification scores. This application can accurately identify promises made during live streams and, through multi-dimensional verification, accurately assess the accuracy of promise information, reducing the false detection rate of promises.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and more specifically, to a method, apparatus, computer equipment, and storage medium for identifying target content in a live streaming room. Background Technology

[0002] With the rapid development of online live-streaming marketing, e-commerce live-streaming has become an important channel for product sales. However, the casual and immediate nature of verbal promises made by live-streamers in live-streaming marketing poses significant challenges to subsequent contract fulfillment management and customer complaint handling.

[0003] Currently, Natural Language Processing (NLP) techniques are typically used to detect sensitive words or identify false advertising in live audio content, or keyword matching is used to determine whether the live content contains prohibited words. However, these methods lack accurate identification of promised information. Furthermore, traditional NLP methods suffer from insufficient recognition capabilities, fragmented output formats of large models, frequent model illusions, and high false recognition recall rates.

[0004] Therefore, there is an urgent need for a method that can accurately identify the promises made in live streams. Summary of the Invention

[0005] In view of this, this application provides a method, apparatus, computer equipment and storage medium for identifying target content in a live broadcast room, which can more accurately identify commitments made during the live broadcast and more accurately evaluate the accuracy of commitment information through multi-dimensional verification, thereby reducing the false detection rate of commitments.

[0006] In a first aspect, embodiments of this application provide a method for identifying target content in a live streaming room, including: Acquire real-time live streaming data from the target car dealer's live streaming room, wherein the real-time live streaming data includes live audio data and live video data; The live audio data, the attribute information of the audio output party, and the session context information are identified and fused to generate role tags; The live audio data is converted into live text data, and the live text data is sliced ​​according to a preset time window. The sliced ​​text data segments are then associated with the role tags, timestamps, and conversation context information to generate structured corpus. Based on the prompt words corresponding to various commitments in the preset prompt word template, the structured corpus is subjected to model recognition to determine the commitment information included in the structured corpus; The commitment information is verified in multiple dimensions to determine a multi-dimensional verification score; wherein the multi-dimensional verification includes: original text anchoring verification for verifying the degree of matching between the commitment information and the corresponding session-based text, context consistency verification for verifying the consistency between the context information at the time of the commitment information output and the commitment information, and business constraint verification for verifying the business rationality of the commitment information. Based on the multi-dimensional verification score, the commitment information is processed for output control.

[0007] In one optional implementation, the step of identifying and fusing the live audio data, the attribute information of the audio output party, and the session context information to generate role tags includes: The voiceprint features of the live audio data are obtained and matched with a pre-stored voiceprint feature database to obtain the recognition result; Obtain the identity information of the account that outputs the live audio; The role label is determined based on the semantic features of the session context information, the identity information, and the recognition result.

[0008] In one optional implementation, the step of performing output control processing on the commitment information based on the multi-dimensional verification score includes: Image recognition is performed on the live video data to extract and display text information; The displayed text information and the commitment information are semantically compared to determine the visual consistency judgment result. Based on the visual consistency discrimination result, an adjustment coefficient is determined; wherein, if the displayed text information matches the commitment information, the adjustment coefficient is a positive number greater than or equal to 1, used to gain the verification score; if the displayed text information conflicts with the commitment information, the adjustment coefficient is a positive number less than 1, used to attenuate the verification score. The verification score is corrected according to the adjustment coefficient to obtain the corrected verification score; Based on the revised verification score, the commitment information is processed for output control.

[0009] In one optional implementation, the commitment information is subjected to output control processing based on the multi-dimensional verification score, including: After determining that the multi-dimensional verification score is greater than the set score threshold, the commitment information is sent to the car dealer user corresponding to the target car dealer's live broadcast room; Receive the target operation submitted by the car dealer user in response to the commitment information, wherein the target operation is one of the following: confirmation operation, modification operation, and supplementation operation; The operation content of the target operation and the commitment information are stored in a database, wherein the data stored in the database is used to perform at least one of the following processes: customer complaint processing, risk tracing processing, model training, and rule optimization.

[0010] In one optional implementation, the commitment information is subjected to output control processing based on the multi-dimensional verification score, including: After determining that the multi-dimensional verification score is greater than the set score threshold, a dynamic standard operating procedure (SOP) rule base for the car dealer live streaming scenario is determined; wherein the dynamic SOP rule base is used to store the attribute information of filtering commitments, and the attribute information of filtering commitments includes the rule information and sample semantic information corresponding to basic service commitments, non-concern commitments and prohibited commitments respectively; Based on a dynamic SOP rule base, rule matching and semantic retrieval are performed on the commitment information to obtain matching results; When the matching result indicates that the commitment information does not match the dynamic SOP rule base, the commitment information is output.

[0011] In one optional implementation, the method further includes: When the matching result indicates that the commitment information matches a basic service commitment or non-concern commitment in the dynamic SOP rule base, the commitment information is filtered. When the matching result indicates that the commitment information matches a prohibited commitment in the dynamic SOP rule base, the commitment information is filtered out, and an alarm message is sent to the target car dealer's live broadcast room.

[0012] In one optional implementation, the commitment information is verified in multiple dimensions to determine a multi-dimensional verification score, including: A target validation algorithm is used to determine the original text support score between the session basis text corresponding to the commitment information and the commitment information. The target validation algorithm includes at least one of the following: longest common subsequence algorithm, edit distance algorithm, and keyword coverage algorithm. Obtain target context corpus data within a preset time window based on the output time of the commitment information, and determine the first semantic vector of the target context corpus data and the second semantic vector of the commitment information; determine the context consistency score based on the first semantic vector and the second semantic vector; Based on a pre-defined business rule knowledge base, the reasonableness of the promised information is verified, and a business credibility score is determined based on the verification results; wherein the pre-defined business rule knowledge base includes: the range of price reduction amount, the range of the number of free gifts, the range of maintenance times, the range of financial services, and the range of after-sales commitments; The multi-dimensional verification score is obtained by weighted summing of the original text support score, the context consistency score, and the business credibility score.

[0013] In one optional implementation, the commitment information is structured data, including: commitment type, commitment content, session basis text, and model confidence level.

[0014] In an optional implementation, after determining the commitment information, the method further includes: The commitment information is standardized to generate processed commitment information, which is used for multi-dimensional verification. The standardization process includes: amount standardization, unit standardization, synonym normalization, and abnormal result filtering. The amount standardization process is used to convert amount data in different expressions into a unified amount format and unit. The synonym normalization process is used to map different words belonging to the same semantic category in the commitment information into unified standard terms based on a preset domain thesaurus. The abnormal result filtering process is used to mark commitment information that cannot be mapped to the preset standard format as abnormal data and discard it.

[0015] In one optional implementation, the step of performing output control processing on the commitment information based on the multi-dimensional verification score includes: After determining that the multi-dimensional verification score is greater than the set score threshold, the historical commitment information of the target car dealer's live broadcast room is obtained from the database; The commitment information is compared with the historical commitment information to obtain the relationship comparison result; When the relationship comparison result indicates a commitment conflict, the commitment information and the historical commitment information are sent to the car dealer user corresponding to the target car dealer's live broadcast room for confirmation, and the commitment information or the historical commitment information is retained based on the confirmation result; When the relationship comparison result indicates commitment coverage, the historical commitment information is deleted and stored in the database; When the relationship comparison result indicates a commitment supplement, the commitment information is merged with the historical commitment information, and the merged commitment information is stored in the database; If the relationship comparison result indicates that the commitment is irrelevant, the commitment information is updated in the database.

[0016] Secondly, embodiments of this application also provide a target content recognition device for a live streaming room, comprising: The acquisition module is used to acquire real-time live streaming data from the target car dealer's live streaming room. The real-time live streaming data includes live audio data and live video data. The first generation module is used to identify and fuse the live audio data, the attribute information of the audio output party, and the session context information to generate role tags; The second generation module is used to convert the live audio data into live text data, slice the live text data according to a preset time window, and associate the sliced ​​text data segments with the role tags, timestamps and conversation context information to generate structured corpus. The first determining module is used to perform model recognition on the structured corpus based on the prompt words corresponding to various commitments in the preset prompt word template, and to determine the commitment information included in the structured corpus. The second determining module is used to perform multi-dimensional verification on the commitment information and determine the multi-dimensional verification score; wherein the multi-dimensional verification includes: original text anchoring verification for verifying the degree of matching between the commitment information and the corresponding session-based text, context consistency verification for verifying the consistency between the context information at the time of the commitment information output and the commitment information, and business constraint verification for verifying the business rationality of the commitment information. The control module is used to perform output control processing on the commitment information based on the multi-dimensional verification score.

[0017] This application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for identifying target content in a live streaming room.

[0018] This application provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the above-mentioned target content recognition method for live streaming rooms.

[0019] The target content recognition method, apparatus, computer equipment, and storage medium provided in this application for live streaming rooms determine the role tags of the audio output party and, after slicing the live streaming text data, associate the sliced ​​text data segments with role tags, timestamps, and conversation context information to generate structured corpus. This structured corpus contains relatively comprehensive and complete information. Then, based on the prompts corresponding to various promises in a preset prompt word template, the model recognizes the structured corpus, accurately determining the promise information included in the structured corpus, and achieving relatively accurate semantic recognition of various promises made by the anchor during the live stream. Furthermore, this application comprehensively verifies the promise information through original text anchoring verification, contextual consistency verification, and business constraint verification. Based on the determined multi-dimensional verification scores, the output of the promise information is controlled. Multi-dimensional verification is used to comprehensively evaluate the authenticity and validity of the promise information, reducing the false recall rate of invalid promises, i.e., reducing the false detection rate of promises. Attached Figure Description

[0020] The accompanying drawings, which are included to provide a further understanding of this specification and form part of this specification, illustrate exemplary embodiments and are used to explain this specification, but do not constitute an undue limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating a target content recognition method for a live streaming room, as shown in an exemplary embodiment of this application; Figure 2 This is a flowchart illustrating a method for identifying target content in a live streaming room, as shown in another exemplary embodiment of this application; Figure 3 This is a schematic diagram of a target content recognition device for a live streaming room, as illustrated in an exemplary embodiment of this application; Figure 4 This is a schematic diagram of the structure of a computer device provided in this application. Detailed Implementation

[0021] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0022] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0023] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0024] To facilitate understanding of this embodiment, a detailed description of the target content recognition method for a live streaming room disclosed in this application embodiment will be provided first. The execution entity of the target content recognition method for a live streaming room provided in this application embodiment is generally a computer device with a certain computing power. This computer device may include, for example, a terminal device, a server, or other processing devices. The terminal device may be a user equipment (UE), a mobile device, a terminal, a personal digital assistant (PDA), a handheld device, a computing device, etc. In some possible implementations, the target content recognition method for a live streaming room can be implemented by a processor calling computer-readable instructions stored in memory.

[0025] See Figure 1 The diagram shows a flowchart of a method for identifying target content in a live streaming room provided in an embodiment of this application. The method includes steps S101 to S106, wherein: S101. Obtain real-time live streaming data from the target car dealer's live streaming room, wherein the real-time live streaming data includes live audio data and live video data; S102. The live audio data, the attribute information of the audio output party, and the session context information are identified and fused to generate role tags; S103. Convert the live audio data into live text data, slice the live text data according to a preset time window, and associate the sliced ​​text data segments with the role tags, timestamps and conversation context information to generate structured corpus. S104. Based on the prompt words corresponding to various commitments in the preset prompt word template, perform model recognition on the structured corpus to determine the commitment information included in the structured corpus; S105. Perform multi-dimensional verification on the commitment information and determine the multi-dimensional verification score; wherein the multi-dimensional verification includes: original text anchoring verification for verifying the degree of matching between the commitment information and the corresponding session-based text, context consistency verification for verifying the consistency between the context information at the time of the commitment information output and the commitment information, and business constraint verification for verifying the business rationality of the commitment information. S106. Based on the multi-dimensional verification score, perform output control processing on the commitment information.

[0026] This application determines the role labels of the audio output party and, after slicing the live text data, associates the sliced ​​text data segments with role labels, timestamps, and conversation context information to generate structured corpus. This structured corpus contains comprehensive and complete information. Then, based on the prompts corresponding to various promises in a preset prompt template, the application performs model recognition on the structured corpus to accurately identify the promise information included in the structured corpus, achieving semantically accurate identification of various promises made by the broadcaster during the live stream. Furthermore, this application comprehensively verifies the promise information through original text anchoring verification, contextual consistency verification, and business constraint verification. Based on the determined multi-dimensional verification scores, the application performs output control processing on the promise information, using multi-dimensional verification to comprehensively evaluate the authenticity and validity of the promise information, reducing the false recall rate of invalid promises and the false detection rate of promises.

[0027] The specific processes of S101-S106 are explained below.

[0028] For S101 and S102, real-time live streaming data from the car dealer's live streaming room is obtained. This real-time live streaming data includes live audio data and live video data.

[0029] It also acquires attribute information of the audio output party, such as identity information, and performs identification and fusion processing on the live audio data, attribute information, and session context information to determine role tags. These role tags may include, for example, car dealers, users, and online sales personnel. The session context information may include audio data generated in the target car dealer's live stream before the current moment and in the historical time period, and / or bullet screen display information.

[0030] Optionally, the step of recognizing and fusing the live audio data, the attribute information of the audio outputter, and the session context information to generate role tags includes: obtaining the voiceprint features of the live audio data and matching them with a pre-stored voiceprint feature library to obtain a recognition result; obtaining the identity information of the account that outputs the live audio; and determining the role tag based on the semantic features of the session context information, the identity information, and the recognition result.

[0031] During implementation, a neural network model can be used to extract the voiceprint features of each audio speaker in the live audio data, and the voiceprint features can be matched with a pre-stored voiceprint feature library to obtain the recognition result. The voiceprint feature library can store the voiceprint features of one or more online sales personnel in the target car dealer's live broadcast room. If the matching score is greater than the preset score, the recognition result is determined to be a successful match. In this case, it means that the voiceprint features belong to the online sales personnel, that is, the probability that their role label is online sales personnel is relatively high.

[0032] For each audio sender, the identity information of the live audio (such as the sender's account) can be obtained. This identity information may include basic account attributes, such as account registration type (belonging to an enterprise or an individual), historical tags (such as whether it has been marked as an online salesperson in the past), etc.

[0033] Furthermore, semantic features of the conversation context can be extracted to determine role labels based on semantic features, recognition results, and identity information. For example, if semantic features indicate that the provider of the live audio has previously introduced vehicle discounts or basic vehicle information, combined with identity information and recognition results (successful match), the role label can be determined to be an online salesperson. Alternatively, if semantic features indicate that the provider of the live audio has previously inquired about vehicle prices and performance, combined with identity information (e.g., an individual user) and recognition results (failed match), the role label can be determined to be a user. Finally, if the identity information is a car dealer's corporate account, the role label is determined to be a car dealer.

[0034] This application uses voiceprint recognition, semantic features of conversation, and account identity information fusion correction to accurately distinguish the identities of different speakers in multi-person live streaming scenarios, thereby accurately labeling the role tag corresponding to each voice message and providing technical assurance for the accurate attribution of commitment information to specific roles.

[0035] For S103 and S104, during implementation, Automatic Speech Recognition (ASR) technology can be used to convert live audio data into live text data, and the live text data can be sliced ​​according to a preset time window. The sliced ​​text data segments are then associated with role tags, timestamps, and conversation context information to generate structured corpus. The length of the preset time window can be set as needed.

[0036] This application can segment a continuous long text stream into short text segments that are easy to process later by slicing it according to a preset time window, while retaining timestamps and role tags, providing structured input data for subsequent commitment recognition.

[0037] The structured corpus is then input into a large language model. The large language model uses prompt words corresponding to various commitments in a preset prompt word template to perform model recognition on the structured corpus and determine the commitment information included in the structured corpus.

[0038] The commitment information is structured data, including: commitment type (promise_type), commitment content (promise_content), session basis text (promise_original_text), and model confidence (promise_confidence).

[0039] The various types of commitments include, but are not limited to: price discount commitments, gift and service commitments, vehicle preparation commitments, financing plan commitments, and after-sales service commitments. The preset prompt templates include prompts corresponding to each type of commitment in a car dealer livestream scenario. These prompts can be extracted from the historical content of each commitment. For example, prompts for price commitments include, but are not limited to, price, price reduction, discount amount, and discount; prompts for after-sales commitments include, but are not limited to, warranty, maintenance, repair, and return / exchange.

[0040] This application uses domain knowledge constraints in a preset prompt word template to constrain the large language model to only recognize promises related to price, gifts, preparation, financial solutions, and after-sales service; it can also force the model to output JSON format results and output the original text reference fields for subsequent credibility verification.

[0041] This application uses preset prompt word templates to guide a large language model to identify various types of commitment information from structured corpora, and can accurately identify various commitments made by the anchor during the live broadcast from a semantic level.

[0042] Optionally, to ensure the accuracy of the verification, after determining the commitment information, the method further includes: standardizing the commitment information to generate processed commitment information, wherein the processed commitment information is used for multi-dimensional verification.

[0043] The standardization process includes: amount standardization, unit standardization, synonym normalization, and abnormal result filtering. The amount standardization process converts amount data in different forms into a unified amount format and unit. The synonym normalization process maps different words belonging to the same semantic category in the commitment information to unified standard terms based on a preset domain thesaurus. The abnormal result filtering process marks commitment information that cannot be mapped to the preset standard format as abnormal data and discards it.

[0044] Amount standardization is used to convert monetary data in different forms into a unified monetary format and unit. For example, expressions such as "3k", "3000 yuan", and "discount of 3,000 yuan" are uniformly converted into "price reduction of 3,000 yuan".

[0045] Standardization of organizational processes, such as converting expressions like "one free maintenance" or "one free maintenance" into a standardized format, such as "one free maintenance".

[0046] When performing synonym normalization, a domain-specific thesaurus can be pre-built to map different words belonging to the same semantic category in the commitment information to unified standard terms, such as mapping words like "instant discount," "preferential treatment," "cheap," and "less charge" to "price reduction."

[0047] Exception filtering can automatically mark results that cannot be mapped to a standard format as anomalous data and discard them to avoid recall of invalid commitments.

[0048] The standardized processing procedure can be set according to the needs of the scenario. For example, it can also include commitment template mapping processing, which maps commitment information to a certain commitment template to generate commitment information in a standard format.

[0049] This application, through the aforementioned standardization process, transforms colloquial and inconsistent commitment statements into a unified, structured standard format, facilitating subsequent multi-dimensional verification, data storage, and statistical analysis.

[0050] For S105, after determining the commitment information, multi-dimensional verification is performed on the commitment information to determine the multi-dimensional verification score, or the processed commitment information is verified in multiple dimensions to determine the multi-dimensional verification score. Multi-dimensional verification includes original text anchoring verification, contextual consistency verification, and business constraint verification. Multi-dimensional verification aims to suppress model illusions caused by large language models and improve the accuracy of commitment detection.

[0051] Optionally, the step of performing multi-dimensional verification on the commitment information and determining a multi-dimensional verification score includes: Step a1: Use a target validation algorithm to determine the original text support score between the session basis text corresponding to the commitment information and the commitment information, wherein the target validation algorithm includes at least one of the following: longest common subsequence algorithm, edit distance algorithm, keyword coverage algorithm; Step a2: Obtain target context corpus data within a preset time window based on the output time of the commitment information, and determine the first semantic vector of the target context corpus data and the second semantic vector of the commitment information; determine the context consistency score based on the first semantic vector and the second semantic vector; Step a3: Based on a preset business rule knowledge base, verify the reasonableness of the commitment information and determine the business credibility score based on the verification results; wherein the preset business rule knowledge base includes: price reduction range, gift quantity range, maintenance frequency range, financial service scope, and after-sales commitment scope; Step a4: The original text support score, the context consistency score, and the business credibility score are weighted and summed to obtain the multi-dimensional verification score.

[0052] In step a1, the target verification algorithm includes at least one of the following: the longest common subsequence algorithm, the edit distance algorithm, and the keyword coverage algorithm. The longest common subsequence algorithm calculates the length of the longest common subsequence between the session-based text and the commitment information, measuring their consistency in character order. The edit distance algorithm calculates the minimum number of edit operations required to convert the session-based text into the commitment information, measuring their literal similarity. The keyword coverage algorithm calculates the hit rate of core keywords in the commitment information within the session-based text, measuring the recall of core semantic elements.

[0053] When the target verification algorithm includes the longest common subsequence algorithm, the edit distance algorithm, and the keyword coverage algorithm, each algorithm can be used to calculate the algorithm score between the session basis text and the commitment information corresponding to the commitment information; the scores of multiple algorithms are weighted and summed to obtain the original text support score.

[0054] In step a2, target contextual data within a preset time window is obtained, based on the output time of the commitment information. For example, the live text within 30 seconds before and after the output time is obtained as target contextual data. Then, a semantic feature extraction network is used to encode the target contextual data into a first semantic vector and the commitment information into a second semantic vector, respectively. The similarity between the two semantic vectors is calculated, and this similarity is determined as the contextual consistency score. The contextual consistency score can be used to determine whether the commitment information conforms to the real conversation context.

[0055] In step a3, a preset business rule knowledge base is queried to verify the reasonableness of the commitment information. The verification includes, but is not limited to, the range of price reductions, the range of free gifts, the range of maintenance services, the range of financial services, and the range of after-sales commitments. A business credibility score is then determined based on the verification results. For example, checking whether the "commitment information: three years of free maintenance" falls within the range of maintenance services; if the reasonableness verification passes, the business credibility score is determined to be 1.0; conversely, if it exceeds the range (i.e., the reasonableness verification fails), a penalty coefficient is calculated based on the degree of exceedance. The greater the exceedance, the lower the business credibility score.

[0056] In step a4, the original text support score, context consistency score, and business credibility score are weighted and summed to obtain a multi-dimensional verification score. The weight coefficients of the original text support score, context consistency score, and business credibility score can be flexibly set according to the needs of the business scenario.

[0057] For S106, after determining the multi-dimensional verification score, if the multi-dimensional verification score is greater than the set score threshold, the commitment information can be output and recorded; the live stream identifier ID, timestamp, original text basis, session context content, credibility, etc. of the commitment information can also be recorded. Conversely, if the multi-dimensional verification score is less than or equal to the score threshold, the commitment information can be deleted without output or recording.

[0058] Optionally, based on the multi-dimensional verification score, the commitment information is processed for output control, including: after determining that the multi-dimensional verification score is greater than a set score threshold, sending the commitment information to the car dealer user corresponding to the target car dealer's live stream; receiving a target operation submitted by the car dealer user for the commitment information, wherein the target operation is one of a confirmation operation, a modification operation, or a supplementary operation; storing the operation content of the target operation and the commitment information in a database, wherein the data stored in the database is used to perform at least one of the following processes: customer complaint handling, risk tracing handling, model training, and rule optimization.

[0059] After confirming that the multi-dimensional verification score exceeds the set score threshold, the commitment information is sent to the corresponding car dealer users in the target car dealer's live stream, such as to the car dealer's management backend or car dealer platform. Car dealer users can manually review and process the commitment information, including confirmation (confirming the accuracy of the commitment information and indicating acceptance), modification (confirming discrepancies in the commitment information, such as incorrect amounts or incomplete descriptions of conditions, and requiring modification), and supplementation (for example, missing additional conditions or supplementary content, and additional limitations).

[0060] The system receives target operations submitted by car dealer users in response to commitment information, and stores the operation content of the target operations and the commitment information in a database. The data stored in the database is used to perform at least one of the following processes: customer complaint handling, risk tracing, model training, and rule optimization.

[0061] In handling customer complaints, when a user files a complaint about a certain promise, the corresponding promise record can be quickly retrieved from the database to verify the authenticity and specific content of the promise.

[0062] Risk tracing can be achieved by tracing the time of creation, output, associated roles, and session context of relevant commitments through a database, so as to use the traced information to handle related commitment disputes.

[0063] Model training involves using manually verified commitment information as positive samples and other filtered commitment information as negative samples. An incremental training dataset is then constructed based on the positive and negative samples, and the large language model is iteratively trained and optimized using the incremental training dataset.

[0064] Rule optimization involves analyzing manually modified and supplemented operation records to identify shortcomings or deficiencies in the dynamic Standard Operating Procedure (SOP) rule base and optimizing its configuration.

[0065] To ensure the accuracy of the committed information, this application may further utilize live video data for visual consistency determination.

[0066] Optionally, the step of performing output control processing on the commitment information based on the multi-dimensional verification score includes: Step b1: Perform image recognition on the live video data and extract the displayed text information; Step b2: Perform a semantic comparison between the displayed text information and the commitment information to determine the visual consistency judgment result; Step b3: Based on the visual consistency discrimination result, determine the adjustment coefficient; wherein, if the displayed text information matches the commitment information, the adjustment coefficient is a positive number greater than or equal to 1, used to gain the verification score; if the displayed text information conflicts with the commitment information, the adjustment coefficient is a positive number less than 1, used to attenuate the verification score. Step b4: Correct the verification score according to the adjustment coefficient to obtain the corrected verification score; Step b5: Based on the corrected verification score, perform output control processing on the commitment information.

[0067] During implementation, image recognition is performed frame-by-frame or frame-by-frame on the live video data, and Optical Character Recognition (OCR) technology is used to extract the text information displayed in the live stream. The displayed text information includes, but is not limited to: the text content of advertisements in the live stream, price tags, product parameter tables, promotional banners, and bullet screen messages.

[0068] The extracted displayed text information is semantically compared with the promised information to determine the visual consistency judgment result. For example, if the host verbally promises "a price reduction of 500 yuan today", and the price tag in the live broadcast shows "a special price reduction of 500 yuan today", then the visual consistency judgment result indicates visual consistency; if the price tag in the image shows "a special price reduction of 50 yuan today", then the visual consistency judgment result indicates visual conflict.

[0069] The adjustment coefficient λ can be determined based on the visual consistency judgment result. For example, if the displayed text information matches the committed information (i.e., visual consistency), the adjustment coefficient λ is determined to be a positive number greater than or equal to 1 (e.g., λ=1.1), which is used to gain the verification score. If the displayed text information conflicts with the committed information (i.e., visual inconsistency), the adjustment coefficient λ is a positive number less than 1 (e.g., λ=0.5), which is used to attenuate the verification score. If no relevant displayed text information is extracted from the live broadcast, the adjustment coefficient λ=1, and no gain or attenuation is applied.

[0070] The multi-dimensional verification score can then be corrected based on the adjustment coefficient λ, for example, the corrected multi-dimensional verification score = original verification score × adjustment coefficient λ. Based on the corrected verification score, the commitment information is then processed for output control. For instance, if the corrected multi-dimensional verification score is greater than a preset score threshold, the commitment information is output and recorded; conversely, if the corrected multi-dimensional verification score is less than or equal to the score threshold, the commitment information is filtered out.

[0071] This application performs image recognition on live video footage, extracts the displayed text information, compares it with the commitment information, obtains the visual consistency judgment result, and then uses the visual consistency judgment result to determine the adjustment coefficient. The adjustment coefficient is then used to correct the gain or attenuation of the verification score. In other words, through visual consistency verification, multimodal cross-verification of audio (such as the anchor's verbal commitment) and video (such as the information displayed on the screen) is achieved, which further improves the accuracy of commitment information verification.

[0072] In another optional implementation, the commitment information is processed for output control based on the multi-dimensional verification score, including: after determining that the multi-dimensional verification score is greater than a set score threshold, determining a dynamic standard operating procedure (SOP) rule base for the car dealer livestream scenario; wherein the dynamic SOP rule base is used to store attribute information for filtering commitments, the attribute information for filtering commitments includes rule information and sample semantic information corresponding to basic service commitments, non-concern commitments, and prohibited commitments, respectively; based on the dynamic SOP rule base, performing rule matching and semantic retrieval on the commitment information to obtain a matching result; when the matching result indicates that the commitment information does not match the dynamic SOP rule base, outputting the commitment information.

[0073] After confirming that the multi-dimensional verification score is greater than the set score threshold, a dynamic SOP rule base for the car dealer live streaming scenario is determined. This rule base stores attribute information for three types of filtering commitments: basic service commitments, non-concern commitments, and prohibited commitments.

[0074] Basic service commitments: such as "one-year warranty" and "platform testing service" are inherent basic services of the platform. These commitments do not need to be specially recorded and can be automatically filtered.

[0075] Non-concerned commitments: Statements such as "fuel consumption description", "vehicle configuration introduction", and "market information description" are subjective sharing or public information. Such content does not constitute a commitment that needs to be fulfilled and can be automatically filtered out.

[0076] Prohibited promises: Promises that violate laws, regulations, or platform rules, such as "full refund," "guaranteed profit," or "lifetime return policy if not satisfied," will be automatically filtered and trigger an alarm.

[0077] The rule information corresponding to each commitment may include keywords extracted from the content of the commitment sample, preset rule content set for filtering commitments, etc.; the sample semantic information may be obtained by extracting semantic features from the content of the commitment sample.

[0078] During implementation, the dynamic SOP rule base supports real-time updates via the ETCD configuration center. For example, when business policies change (such as adding prohibited commitment types or adjusting the scope of basic services), operations personnel only need to modify the corresponding configuration items in the ETCD configuration center. The system executing this method uses ETCD's Watch mechanism to detect configuration changes in real time and automatically load the new rules without restarting the service. This allows the system to quickly adapt to frequent changes in live-streaming marketing policies and load the latest rules without restarting the service, significantly improving the system's flexibility and maintainability.

[0079] Furthermore, based on the dynamic SOP rule base, the commitment information can be subjected to dual matching, namely rule matching and semantic retrieval, to obtain the matching results. Rule matching involves matching the commitment information against the structured rules in the dynamic SOP rule base (such as keyword matching, regular expression matching, Boolean logic matching, etc.) to quickly determine whether the commitment information belongs to one of the three types of filtering commitments.

[0080] For commitments that do not match the rules, the commitment information can be vectorized. The resulting semantic vector is then compared with the sample semantic information in the dynamic SOP rule base to determine whether it is semantically similar to the filtered commitment. For example, when a broadcaster says "I'll take care of your engine oil for life," although it does not match the keyword "lifetime free maintenance," it is highly similar to the prohibited commitment semantically. Through semantic search, it can be determined that it matches the prohibited commitment and belongs to the commitment that should be filtered.

[0081] If the matching result indicates that the commitment information does not match the dynamic SOP rule base (i.e., it does not belong to any of the basic service commitment, non-concern commitment, and prohibited commitment), then the commitment information will be output.

[0082] Optionally, the method further includes: when the matching result indicates that the commitment information matches the basic service commitment or non-concern commitment in the dynamic SOP rule base, determining to filter the commitment information, that is, not outputting or recording it.

[0083] When the matching result indicates that the commitment information matches the prohibited commitment in the dynamic SOP rule base, the commitment information is filtered, and an alarm message is sent to the target car dealer's live broadcast room to remind the operators to pay attention to the host's violation of the rules.

[0084] This application sets up three types of filtering commitments: basic service commitments, non-concern commitments, and prohibited commitments. Each type uses a different filtering strategy and is stored in a dynamic SOP rule base. By using the dynamic SOP rule base to filter commitment information, it can filter out invalid chat content and issue alerts for prohibited commitments, thus achieving refined management of commitment information.

[0085] Optionally, the step of output control processing of the commitment information based on the multi-dimensional verification score includes: after determining that the multi-dimensional verification score is greater than a set score threshold, obtaining historical commitment information of the target car dealer's live stream from the database; comparing the commitment information with the historical commitment information to obtain a relationship comparison result; when the relationship comparison result indicates a commitment conflict, sending the commitment information and the historical commitment information to the car dealer user corresponding to the target car dealer's live stream for confirmation, and retaining the commitment information or the historical commitment information according to the confirmation result; when the relationship comparison result indicates a commitment overriding, deleting the historical commitment information and storing the commitment information in the database; when the relationship comparison result indicates a commitment supplement, merging the commitment information with the historical commitment information and storing the merged commitment information in the database; and when the relationship comparison result indicates a commitment irrelevant, updating the commitment information in the database.

[0086] After determining that the multi-dimensional verification score is greater than the set score threshold, the currently detected commitment information can be compared with the historical commitment information to determine the relationship between the commitment information and the historical commitment information.

[0087] For example, for each historical commitment, based on the commitment information and the content of the historical commitment information, it is determined whether they target the same object and whether the commitment types are the same. If the target objects are different or the commitment types are different, then the two commitments are unrelated. If they target the same object (i.e., the same configuration and the same model of vehicle) and the commitment types are the same, then it is further checked whether the content of the two commitments is related. If the commitment content is on the same dimension and logically mutually exclusive, then the commitment content conflicts. If the commitment content is on the same dimension and the current commitment is more favorable than the historical commitment, then the commitment content overlaps. If the current commitment is a supplement to the historical commitment, then the commitment content is supplementary.

[0088] For example, if the historical promise information states that vehicle A is not negotiable, and the current promise information states that vehicle A offers a discount of 2000 yuan, then the comparison result is a promise conflict. If the historical promise information states that vehicle A offers a discount of 1000 yuan, and the current promise information states that vehicle A offers a discount of 3000 yuan, then the comparison result is a promise coverage. For another example, if the historical promise information states that vehicle A enjoys a three-year warranty, and the current promise information states that vehicle A enjoys a three-year warranty when it is a non-commercial vehicle, then the current promise information can be considered a supplement to the historical promise information. If the historical promise information states that ordering vehicle A today includes free floor mats and window tinting, and the current promise information states that ordering vehicle A today also includes a free dashcam, then the current promise information is a supplement to the historical promise information. Optionally, the comparison result can also include conditional dependencies. For example, if the current promise depends on the preceding state of the historical promise, then the promise content has a conditional dependency; for example, the historical promise: vehicle A is priced at a special price of 130,000 yuan; the current promise: pay a deposit today to reserve, and receive a discount of 3000 yuan on top of the special price. When the relationship comparison result is a conditional dependency, the current commitment information can be associated with historical commitment information and stored in the database.

[0089] When the relationship comparison result indicates a conflict of commitments, in order to ensure the accuracy of the commitments, the commitment information and historical commitment information can be sent to the car dealer users in the target car dealer's live broadcast room for confirmation, and the commitment information or historical commitment information can be retained based on the confirmation result; at the same time, a prompt message can also be sent to the target car dealer's live broadcast room to prompt the online sales personnel to explain the commitment information.

[0090] When the relation comparison result is a commitment overwrite, the current commitment information is used to overwrite the historical commitment information, that is, the historical commitment information is deleted and the commitment information is stored in the database.

[0091] When the relationship comparison result indicates that the commitment is supplementary, the commitment information is merged with the historical commitment information, and the merged commitment information is stored in the database; when the relationship comparison result indicates that the commitment is irrelevant, the commitment information is updated in the database.

[0092] The above process allows for the joint verification and correction of multiple promises made by the target car dealer's livestream during the livestream, ensuring the accuracy of these promises.

[0093] Combination Figure 2 The logical diagram shown illustrates the target content identification method for live streaming rooms proposed in this application.

[0094] In the data acquisition layer, real-time live data from the target car dealer's live stream is obtained, including live audio data and live video data.

[0095] In the data processing layer, voiceprint features are extracted from live audio data, and the semantic features of the conversation context information, identity information, and voiceprint feature recognition results are fused and corrected to determine role labels. Additionally, the converted live text data can be sliced, and the sliced ​​text data segments are associated with role labels, timestamps, and conversation context information to generate structured corpus.

[0096] In the commitment recognition layer, a large language model is used to identify the commitment information in the structured corpus based on the prompt words corresponding to various commitments in the preset prompt word template. The commitment information is then standardized, such as amount standardization, unit standardization, synonym normalization, and abnormal result filtering.

[0097] In the multi-dimensional verification layer, the commitment information is verified in multiple dimensions to obtain the original text support score, context consistency score, and business credibility score. The original text support score, context consistency score, and business credibility score are then weighted and summed to obtain the multi-dimensional verification score.

[0098] In the output control layer, it can be determined whether the multi-dimensional verification score is greater than the set score threshold. If so, output control processing can be performed, such as manual review and storage processing, SOP rule filtering processing, visual consistency verification processing, and historical commitment relationship verification. If not, the commitment information is filtered, that is, it is not processed or recorded.

[0099] The manual review and storage process includes: sending the commitment information to the car dealer users corresponding to the target car dealer's live stream; receiving the target operation submitted by the car dealer users regarding the commitment information, where the target operation is one of the following: confirmation operation, modification operation, or supplementary operation; storing the operation content of the target operation and the commitment information in the database, where the data stored in the database is used to perform at least one of the following processes: customer complaint handling, risk tracing handling, model training, and rule optimization. That is, the data stored in the database can be used for rule optimization, such as updating the dynamic SOP rule base so that the updated SOP rule base can be used to perform the SOP rule filtering process later; or, the mapping relationship of the adjustment coefficient of visual consistency verification can be adjusted, etc.

[0100] For example, incremental training samples can be built based on data in the database, and the large language model can be iteratively trained and optimized using the incremental training samples so that the optimized large language model can be used for the commitment recognition process in the future; or, the preset prompt word template can be updated using data in the database so that the updated prompt word template can be used to more accurately identify commitment information.

[0101] During implementation, customer complaint processing and risk tracing can be performed using data from the database, and then the processing results can be generated. The processing results can then be used to perform model training and rule optimization.

[0102] SOP rule filtering processing includes: determining a dynamic standard operating procedure (SOP) rule base for the car dealer livestream scenario; the dynamic SOP rule base stores the attribute information of filtering commitments, including rule information and sample semantic information corresponding to basic service commitments, non-concern commitments, and prohibited commitments; based on the dynamic SOP rule base, rule matching and semantic retrieval are performed on the commitment information to obtain matching results; when the matching result indicates a mismatch, the commitment information is output; when the matching result indicates that the commitment information matches the basic service commitment or non-concern commitment, the filtering commitment information is determined; when the matching result indicates that the commitment information matches the prohibited commitment, the filtering commitment information is determined, and an alarm message is sent to the target car dealer livestream room.

[0103] The visual consistency verification process includes: performing image recognition on the live video data to extract the displayed text information; performing semantic comparison between the displayed text information and the commitment information to determine the visual consistency discrimination result; determining the adjustment coefficient based on the visual consistency discrimination result; for example, if the displayed text information matches the commitment information, then the adjustment coefficient λ≥1, and the verification score is increased; if the displayed text information conflicts with the commitment information, then the adjustment coefficient λ<1, and the verification score is decreased; if there is no information, then the adjustment coefficient λ=1; then the verification score can be corrected using the adjustment coefficient λ to obtain the corrected verification score; and output control processing of the commitment information is performed based on the corrected verification score.

[0104] Historical commitment relationship verification includes comparing current commitment information with historical commitment information to determine the comparison result, allowing for appropriate processing of the commitment information based on the result. For example, if the comparison result indicates a commitment conflict, manual review of the current and historical commitment information can be used to determine whether to retain the current or historical commitment information. Another example is if the comparison result indicates commitment overriding, using the current commitment information to overwrite the historical commitment information, retaining only the latest commitment information. Yet another example is if the comparison result indicates commitment supplementation, merging the current and historical commitment information and storing the merged commitment information in a database.

[0105] In practice, one of the following processes can be executed: manual review and storage, SOP rule filtering, visual consistency verification, and historical commitment relationship verification. Alternatively, multiple processes can be executed in sequence. The combination and order of these processes can be configured according to business requirements; the following is merely an example.

[0106] For example, manual review and storage processing, SOP rule filtering processing, and visual consistency verification processing can be performed. Specifically, after generating multi-dimensional verification scores (or after determining that the multi-dimensional verification scores are greater than the set score threshold), visual consistency verification processing is performed, that is, the multi-dimensional verification scores are corrected to obtain corrected scores, and then the corrected scores are compared with the threshold. If it is determined that the corrected scores are greater than the threshold, the SOP rule filtering process is performed, using a dynamic SOP rule base to filter the commitment information; if it is determined that the matching result is a mismatch, manual review and storage operations are performed.

[0107] For example, after generating multi-dimensional verification scores, visual consistency verification can be performed to obtain corrected scores. If the corrected scores are greater than a threshold, historical commitment relationship verification can be performed.

[0108] This application, through a full-process processing including identification, standardization, verification, filtering, and confirmation, enables the accurate identification and standardized output of personalized commitment information such as price discounts, free gifts and services, vehicle preparation, and financial solutions in car dealer live streaming scenarios. This improves the accuracy, credibility, and quality inspection efficiency of commitment identification, avoids interference from invalid information, and provides reliable data support for customer complaint tracing, risk control, and model optimization.

[0109] This application constructs an end-to-end closed-loop processing logic through the aforementioned process, from real-time online identification to offline data feedback, and then from offline optimization to feed back into online strategies. Through this process, this application can alleviate the problem of continuously declining accuracy in traditional static models and fixed rule bases. Specifically, the closed loop formed by data feedback and model / rule optimization allows the system to continuously learn from actual business feedback. Each manual review confirmation result, each customer complaint handling result, or each modification of a car dealer's commitment data contributes to the system's experience. Furthermore, as the system's runtime increases, the accuracy of commitment recognition in the target car dealer's live-streaming scenario improves, while the recall rate of invalid commitments decreases, achieving adaptive evolution of the system.

[0110] The closed-loop mechanism of this application can provide an engineered solution for hallucination suppression. When the model outputs commitment information that deviates from the facts due to hallucination, the erroneous result will be identified and corrected through a manual review process. The corrected manually labeled data will be fed back to the training set as a negative feedback signal and used directly to suppress the model's tendency to hallucinate on this type of problem in the next iteration, thus forming a continuous correction and suppression of the model's hallucination.

[0111] This application also enables dynamic adaptive updates to the business rule base. Considering the frequent changes in marketing policies for car dealer livestreams, a static rule base is difficult to adapt to. Therefore, this application uses the ETCD configuration center to achieve real-time updates of SOP rules, and can also analyze offline data to identify blind spots in the rule base's coverage. For example, when a new expression repeatedly appears and is modified by car dealer users, the system can identify this pattern and generate rule supplement suggestions, ensuring the rule base remains synchronized with real business scenarios and avoiding the problem of rules lagging behind business changes.

[0112] This application constructs a positive feedback loop of "data-knowledge-strategy," transforming raw business data (audio and video streams) into structured knowledge (standardized commitment information), then elevating that knowledge to optimization strategies (model parameter updates, rule configuration adjustments) through human feedback. Finally, these strategies feed back into the online system to improve the accuracy of the next round of data processing. This positive feedback loop achieves in-depth mining and reuse of data value at the system level, ensuring that each instance of live data processing lays the foundation for the intelligent upgrade of the entire system.

[0113] This application establishes a traceable chain of decision-making evidence, encompassing data from all stages of the closed-loop process, including the original ASR text, model recognition results, verification scores for each dimension, SOP rule compliance, and manual review records, all of which can be stored in a structured database. When a customer complaint or risk event occurs, the system can completely trace the entire process of a commitment from its inception to its output, clearly identifying the point of origin of the problem (whether it's model recognition bias, improper verification threshold settings, or incorrect SOP rule configuration), providing data support for accurate problem localization and targeted optimization.

[0114] Corresponding to the aforementioned embodiments of the target content recognition method for live streaming rooms, this application also provides embodiments of the target content recognition device for live streaming rooms. Figure 3 A schematic diagram of the target content recognition device for a live streaming room provided in this application specifically includes: The acquisition module 301 is used to acquire real-time live streaming data from the target car dealer's live streaming room, wherein the real-time live streaming data includes live audio data and live video data. The first generation module 302 is used to identify and fuse the live audio data, the attribute information of the audio output party, and the session context information to generate role tags; The second generation module 303 is used to convert the live audio data into live text data, slice the live text data according to a preset time window, and associate the sliced ​​text data segments with the role tags, timestamps and conversation context information to generate structured corpus. The first determining module 304 is used to perform model recognition on the structured corpus based on the prompt words corresponding to various commitments in the preset prompt word template, and to determine the commitment information included in the structured corpus. The second determining module 305 is used to perform multi-dimensional verification on the commitment information and determine the multi-dimensional verification score; wherein the multi-dimensional verification includes: original text anchoring verification for verifying the degree of matching between the commitment information and the corresponding session basis text, context consistency verification for verifying the consistency between the context information at the time of output of the commitment information and the commitment information, and business constraint verification for verifying the business rationality of the commitment information. The control module 306 is used to perform output control processing on the commitment information based on the multi-dimensional verification score.

[0115] In one optional implementation, when the first generation module 302 performs identification and fusion processing on the live audio data, the attribute information of the audio output party, and the session context information to generate a role tag, it is used to: obtain the voiceprint features of the live audio data and match them with a pre-stored voiceprint feature library to obtain an identification result; obtain the identity information of the account that outputs the live audio; and determine the role tag based on the semantic features of the session context information, the identity information, and the identification result.

[0116] In one optional implementation, when the control module 306 performs output control processing on the commitment information based on the multi-dimensional verification score, it is used to: Image recognition is performed on the live video data to extract and display text information; The displayed text information and the commitment information are semantically compared to determine the visual consistency judgment result. Based on the visual consistency discrimination result, an adjustment coefficient is determined; wherein, if the displayed text information matches the commitment information, the adjustment coefficient is a positive number greater than or equal to 1, used to gain the verification score; if the displayed text information conflicts with the commitment information, the adjustment coefficient is a positive number less than 1, used to attenuate the verification score. The verification score is corrected according to the adjustment coefficient to obtain the corrected verification score; Based on the revised verification score, the commitment information is processed for output control.

[0117] In one optional implementation, when the control module 306 performs output control processing on the commitment information based on the multi-dimensional verification score, it is used to: After determining that the multi-dimensional verification score is greater than the set score threshold, the commitment information is sent to the car dealer user corresponding to the target car dealer's live broadcast room; Receive the target operation submitted by the car dealer user in response to the commitment information, wherein the target operation is one of the following: confirmation operation, modification operation, and supplementation operation; The operation content of the target operation and the commitment information are stored in a database, wherein the data stored in the database is used to perform at least one of the following processes: customer complaint processing, risk tracing processing, model training, and rule optimization.

[0118] In one optional implementation, when the control module 306 performs output control processing on the commitment information based on the multi-dimensional verification score, it is used to: After determining that the multi-dimensional verification score is greater than the set score threshold, a dynamic standard operating procedure (SOP) rule base for the car dealer live streaming scenario is determined; wherein the dynamic SOP rule base is used to store the attribute information of filtering commitments, and the attribute information of filtering commitments includes the rule information and sample semantic information corresponding to basic service commitments, non-concern commitments and prohibited commitments respectively; Based on a dynamic SOP rule base, rule matching and semantic retrieval are performed on the commitment information to obtain matching results; When the matching result indicates that the commitment information does not match the dynamic SOP rule base, the commitment information is output.

[0119] In an optional implementation, the control module 306 is further configured to: When the matching result indicates that the commitment information matches a basic service commitment or non-concern commitment in the dynamic SOP rule base, the commitment information is filtered. When the matching result indicates that the commitment information matches a prohibited commitment in the dynamic SOP rule base, the commitment information is filtered out, and an alarm message is sent to the target car dealer's live broadcast room.

[0120] In one optional implementation, when the second determining module 305 performs multi-dimensional verification on the commitment information and determines the multi-dimensional verification score, it is used to: A target validation algorithm is used to determine the original text support score between the session basis text corresponding to the commitment information and the commitment information. The target validation algorithm includes at least one of the following: longest common subsequence algorithm, edit distance algorithm, and keyword coverage algorithm. Obtain target context corpus data within a preset time window based on the output time of the commitment information, and determine the first semantic vector of the target context corpus data and the second semantic vector of the commitment information; determine the context consistency score based on the first semantic vector and the second semantic vector; Based on a pre-defined business rule knowledge base, the reasonableness of the promised information is verified, and a business credibility score is determined based on the verification results; wherein the pre-defined business rule knowledge base includes: the range of price reduction amount, the range of the number of free gifts, the range of maintenance times, the range of financial services, and the range of after-sales commitments; The multi-dimensional verification score is obtained by weighted summing of the original text support score, the context consistency score, and the business credibility score.

[0121] In one optional implementation, the commitment information is structured data, including: commitment type, commitment content, session basis text, and model confidence level.

[0122] In an optional implementation, after determining the commitment information, the first determining module 304 is further configured to: The commitment information is standardized to generate processed commitment information, which is used for multi-dimensional verification. The standardization process includes: amount standardization, unit standardization, synonym normalization, and abnormal result filtering. The amount standardization process is used to convert amount data in different expressions into a unified amount format and unit. The synonym normalization process is used to map different words belonging to the same semantic category in the commitment information into unified standard terms based on a preset domain thesaurus. The abnormal result filtering process is used to mark commitment information that cannot be mapped to the preset standard format as abnormal data and discard it.

[0123] In one optional implementation, when the control module 306 performs output control processing on the commitment information based on the multi-dimensional verification score, it is used to: After determining that the multi-dimensional verification score is greater than the set score threshold, the historical commitment information of the target car dealer's live broadcast room is obtained from the database; The commitment information is compared with the historical commitment information to obtain the relationship comparison result; When the relationship comparison result indicates a commitment conflict, the commitment information and the historical commitment information are sent to the car dealer user corresponding to the target car dealer's live broadcast room for confirmation, and the commitment information or the historical commitment information is retained based on the confirmation result; When the relationship comparison result indicates commitment coverage, the historical commitment information is deleted and stored in the database; When the relationship comparison result indicates a commitment supplement, the commitment information is merged with the historical commitment information, and the merged commitment information is stored in the database; If the relationship comparison result indicates that the commitment is irrelevant, the commitment information is updated in the database.

[0124] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0125] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and 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 modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0126] This application also provides a computer-readable storage medium storing a computer program that can be used to execute the target content recognition method for live streaming rooms described in the above embodiments.

[0127] This application also provides a computer device, see [link to relevant documentation] Figure 4 The diagram shows the structure of the computer device provided in this application. At the hardware level, the computer device includes a processor, an internal bus, a network interface, memory, and non-volatile memory, and may also include other hardware required for business operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then runs it to implement the target content recognition method for the live streaming room described in the above embodiments. Of course, in addition to software implementation, this specification does not exclude other implementation methods, such as logic devices or a combination of hardware and software, etc. That is to say, the execution subject of the following processing flow is not limited to each logic unit, but can also be hardware or logic devices.

[0128] Those skilled in the art will understand that embodiments of this application can be provided as methods, apparatus, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs) containing computer-usable program code. The form of a computer program product implemented on ROM, optical memory, etc.

[0129] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0130] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0131] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the functions specified in one or more boxes. In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0132] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0133] Computer-readable media, including both permanent and non-permanent, removable and non-removable media, can store information using any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, and optical disc read-only memory (CD). ROM, digital multifunction optical disc (DVD) or other optical storage, magnetic cassette tape, magnetic disk storage or other magnetic storage devices or any other non-transfer media, may be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0134] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0135] This specification can be described in the general context of computer-executable instructions that are executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This specification can also be practiced in distributed computing environments, where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0136] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0137] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for identifying target content in a live streaming room, characterized in that, The method includes: Acquire real-time live streaming data from the target car dealer's live streaming room, wherein the real-time live streaming data includes live audio data and live video data; The live audio data, the attribute information of the audio output party, and the session context information are identified and fused to generate role tags; The live audio data is converted into live text data, and the live text data is sliced ​​according to a preset time window. The sliced ​​text data segments are then associated with the role tags, timestamps, and conversation context information to generate structured corpus. Based on the prompt words corresponding to various commitments in the preset prompt word template, the structured corpus is subjected to model recognition to determine the commitment information included in the structured corpus; The commitment information is verified in multiple dimensions to determine a multi-dimensional verification score; wherein the multi-dimensional verification includes: original text anchoring verification for verifying the degree of matching between the commitment information and the corresponding session-based text, context consistency verification for verifying the consistency between the context information at the time of the commitment information output and the commitment information, and business constraint verification for verifying the business rationality of the commitment information. Based on the multi-dimensional verification score, the commitment information is processed for output control.

2. The method according to claim 1, characterized in that, The process of identifying and fusing the live audio data, the attribute information of the audio output party, and the session context information to generate role tags includes: The voiceprint features of the live audio data are obtained and matched with a pre-stored voiceprint feature database to obtain the recognition result; Obtain the identity information of the account that outputs the live audio; The role label is determined based on the semantic features of the session context information, the identity information, and the recognition result.

3. The method according to claim 1, characterized in that, The step of output control processing of the commitment information based on the multi-dimensional verification score includes: Image recognition is performed on the live video data to extract and display text information; The displayed text information and the commitment information are semantically compared to determine the visual consistency judgment result. Based on the visual consistency discrimination result, an adjustment coefficient is determined; wherein, if the displayed text information matches the commitment information, the adjustment coefficient is a positive number greater than or equal to 1, used to gain the verification score; if the displayed text information conflicts with the commitment information, the adjustment coefficient is a positive number less than 1, used to attenuate the verification score. The verification score is corrected according to the adjustment coefficient to obtain the corrected verification score; Based on the revised verification score, the commitment information is processed for output control.

4. The method according to claim 1, characterized in that, Based on the multi-dimensional verification score, the commitment information is subjected to output control processing, including: After determining that the multi-dimensional verification score is greater than the set score threshold, the commitment information is sent to the car dealer user corresponding to the target car dealer's live broadcast room; Receive the target operation submitted by the car dealer user in response to the commitment information, wherein the target operation is one of the following: confirmation operation, modification operation, and supplementation operation; The operation content of the target operation and the commitment information are stored in a database, wherein the data stored in the database is used to perform at least one of the following processes: customer complaint processing, risk tracing processing, model training, and rule optimization.

5. The method according to claim 1, characterized in that, Based on the multi-dimensional verification score, the commitment information is subjected to output control processing, including: After determining that the multi-dimensional verification score is greater than the set score threshold, a dynamic standard operating procedure (SOP) rule base for the car dealer live streaming scenario is determined; wherein the dynamic SOP rule base is used to store the attribute information of filtering commitments, and the attribute information of filtering commitments includes the rule information and sample semantic information corresponding to basic service commitments, non-concern commitments and prohibited commitments respectively; Based on a dynamic SOP rule base, rule matching and semantic retrieval are performed on the commitment information to obtain matching results; When the matching result indicates that the commitment information does not match the dynamic SOP rule base, the commitment information is output.

6. The method according to claim 5, characterized in that, The method further includes: When the matching result indicates that the commitment information matches a basic service commitment or non-concern commitment in the dynamic SOP rule base, the commitment information is filtered. When the matching result indicates that the commitment information matches a prohibited commitment in the dynamic SOP rule base, the commitment information is filtered out, and an alarm message is sent to the target car dealer's live broadcast room.

7. The method according to any one of claims 1-6, characterized in that, The process of performing multi-dimensional verification on the commitment information and determining the multi-dimensional verification score includes: A target validation algorithm is used to determine the original text support score between the session basis text corresponding to the commitment information and the commitment information. The target validation algorithm includes at least one of the following: longest common subsequence algorithm, edit distance algorithm, and keyword coverage algorithm. Obtain target context corpus data within a preset time window based on the output time of the commitment information, and determine the first semantic vector of the target context corpus data and the second semantic vector of the commitment information; determine the context consistency score based on the first semantic vector and the second semantic vector; Based on a pre-defined business rule knowledge base, the reasonableness of the promised information is verified, and a business credibility score is determined based on the verification results; wherein the pre-defined business rule knowledge base includes: the range of price reduction amount, the range of the number of free gifts, the range of maintenance times, the range of financial services, and the range of after-sales commitments; The multi-dimensional verification score is obtained by weighted summing of the original text support score, the context consistency score, and the business credibility score.

8. The method according to any one of claims 1-6, characterized in that, The commitment information is structured data, including: commitment type, commitment content, conversation basis text, and model confidence level.

9. The method according to any one of claims 1-6, characterized in that, After determining the commitment information, the method further includes: The commitment information is standardized to generate processed commitment information, which is used for multi-dimensional verification. The standardization process includes: amount standardization, unit standardization, synonym normalization, and abnormal result filtering. The amount standardization process is used to convert amount data in different expressions into a unified amount format and unit. The synonym normalization process is used to map different words belonging to the same semantic category in the commitment information into unified standard terms based on a preset domain thesaurus. The abnormal result filtering process is used to mark commitment information that cannot be mapped to the preset standard format as abnormal data and discard it.

10. The method according to any one of claims 1-6, characterized in that, The step of output control processing of the commitment information based on the multi-dimensional verification score includes: After determining that the multi-dimensional verification score is greater than the set score threshold, the historical commitment information of the target car dealer's live broadcast room is obtained from the database; The commitment information is compared with the historical commitment information to obtain the relationship comparison result; When the relationship comparison result indicates a commitment conflict, the commitment information and the historical commitment information are sent to the car dealer user corresponding to the target car dealer's live broadcast room for confirmation, and the commitment information or the historical commitment information is retained based on the confirmation result; When the relationship comparison result indicates commitment coverage, the historical commitment information is deleted and stored in the database; When the relationship comparison result indicates a commitment supplement, the commitment information is merged with the historical commitment information, and the merged commitment information is stored in the database; If the relationship comparison result indicates that the commitment is irrelevant, the commitment information is updated in the database.

11. A target content recognition device for a live streaming room, characterized in that, The device includes: The acquisition module is used to acquire real-time live streaming data from the target car dealer's live streaming room. The real-time live streaming data includes live audio data and live video data. The first generation module is used to identify and fuse the live audio data, the attribute information of the audio output party, and the session context information to generate role tags; The second generation module is used to convert the live audio data into live text data, slice the live text data according to a preset time window, and associate the sliced ​​text data segments with the role tags, timestamps and conversation context information to generate structured corpus. The first determining module is used to perform model recognition on the structured corpus based on the prompt words corresponding to various commitments in the preset prompt word template, and to determine the commitment information included in the structured corpus. The second determining module is used to perform multi-dimensional verification on the commitment information and determine the multi-dimensional verification score; wherein the multi-dimensional verification includes: original text anchoring verification for verifying the degree of matching between the commitment information and the corresponding session-based text, context consistency verification for verifying the consistency between the context information at the time of the commitment information output and the commitment information, and business constraint verification for verifying the business rationality of the commitment information. The control module is used to perform output control processing on the commitment information based on the multi-dimensional verification score.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by a processor, it implements the steps of the method according to any one of claims 1-10.

13. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor performs the steps of the method according to any one of claims 1-10.