Data processing method and device, large model fine-tuning method and device

Abstract

The present disclosure provides a data processing method and device, relating to the fields of computer technology, specifically to large models, deep learning, natural language processing, and other technical fields. The specific implementation scheme comprises: obtaining interactive message data and behavioral operation data of service recipients on a service platform; obtaining sentiment assessment results and semantic sentiment labels based on the interactive message data and a content understanding large model; determining the behavioral trajectory information of the service recipient based on the semantic sentiment labels and the behavioral operation data; and determining the behavioral assessment results of the service recipient based on the sentiment assessment results, the behavioral operation data, and the behavioral trajectory information, thereby improving the accuracy of risky user detection.

Description

Technical Field

[0001] The present disclosure relates to the field of computers, specifically to technical fields such as large models, deep learning, and natural language processing, and in particular to a data processing method and device, a large model fine-tuning method and device, an electronic device, a computer-readable storage medium, and a computer program product. Background Art

[0002] In the network service scenario, existing technologies generally rely on sensitive word libraries combined with simple text review rules to capture sensitive words used abnormally by service recipients in comments.

[0003] However, the above-mentioned sensitive word monitoring method that relies on vocabulary matching cannot capture the highly variable abnormal language of the service recipients. These fixed text review rules are very likely to miss complex network terms or expressions with implicit offensive meanings. The words and behaviors of the service recipients may have adverse effects on the service objects on the service platform. Summary of the Invention

[0004] The present disclosure provides a data processing method and apparatus, a large model fine-tuning method and apparatus, an electronic device, a computer-readable storage medium, and a computer program product.

[0005] According to the first aspect, a data processing method is provided, which includes: obtaining interactive message data and behavioral operation data of a service object in a service platform; obtaining sentiment evaluation results and semantic sentiment labels based on the interactive message data and a content understanding model; determining behavioral trajectory information of the service object based on the semantic sentiment labels and behavioral operation data; determining the behavioral evaluation results of the service object based on the sentiment evaluation results, behavioral operation data and behavioral trajectory information.

[0006] According to the second aspect, a large model fine-tuning method is provided, which includes: obtaining a training sample set and a pre-trained large model, the training sample set includes at least one training sample, and the training sample includes: sample message data and multi-task annotation data corresponding to the sample message data; selecting training samples from the training sample set; inputting the selected training samples into the large model to obtain the sentiment evaluation results and semantic sentiment labels output by the large model; training the large model based on the sentiment evaluation results, semantic sentiment labels and the selected training samples to obtain a trained content understanding large model.

[0007] According to the third aspect, a data processing device is provided, which includes: a data acquisition unit, configured to acquire interactive message data and behavioral operation data of a service object in a service platform; an acquisition unit, configured to obtain sentiment evaluation results and semantic sentiment labels based on the interactive message data and a content understanding model; a trajectory determination unit, configured to determine behavioral trajectory information of the service object based on the semantic sentiment labels and behavioral operation data; and a result determination unit, configured to determine the behavioral evaluation result of the service object based on the sentiment evaluation result, behavioral operation data, and behavioral trajectory information.

[0008] According to the fourth aspect, a large model fine-tuning device is provided, which includes: an information acquisition unit, configured to acquire a training sample set and a pre-trained large model, the training sample set includes at least one training sample, and the training sample includes: sample message data and multi-task annotation data corresponding to the sample message data; a selection unit, configured to select training samples from the training sample set; an input unit, configured to input the selected training samples into the large model, and obtain the sentiment evaluation results and semantic sentiment labels output by the large model; a training unit, configured to train the large model based on the sentiment evaluation results, semantic sentiment labels and selected training samples, and obtain a trained content understanding large model.

[0009] According to the fifth aspect, an electronic device is provided, which includes: at least one processor; and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the method described in any implementation of the first aspect or the second aspect.

[0010] According to a sixth aspect, a non-transitory computer-readable storage medium storing computer instructions is provided, where the computer instructions are used to cause a computer to execute the method described in any implementation of the first aspect or the second aspect.

[0011] According to a seventh aspect, a computer program product is provided, comprising a computer program, which, when executed by a processor, implements the method described in any one of the implementations of the first aspect or the second aspect.

[0012] The data processing method and device provided by the embodiments of the present disclosure first obtain the interactive message data and behavioral operation data of the service object in the service platform; secondly, based on the interactive message data and the content understanding large model, obtain the sentiment evaluation results and semantic sentiment labels; thirdly, based on the semantic sentiment labels and behavioral operation data, determine the behavioral trajectory information of the service object; finally, based on the sentiment evaluation results, behavioral operation data and behavioral trajectory information, determine the behavioral evaluation results of the service object. Thus, not only the sentiment analysis is performed on the interactive message data of the service object, but also the behavioral operation data of the service object on the service platform is combined, and the large model is used for global semantics and sentiment understanding of the context. This method can handle complex and ambiguous attack methods in the service platform, locate high-risk users, improve the accuracy of risk user detection, and reduce the damage of malicious attacks to the service platform ecology.

[0013] It should be understood that the contents described in this section are not intended to identify the key or important features of the embodiments of the present disclosure, nor are they intended to limit the scope of the present disclosure. Other features of the present disclosure will become readily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] The accompanying drawings are provided to facilitate a better understanding of the present invention and do not constitute a limitation of the present disclosure.

[0015] Figure 1 is a flow chart of an embodiment of a data processing method according to the present disclosure;

[0016] Figure 2 This is a structural diagram of data interaction among various modules in the data processing process of the present disclosure;

[0017] Figure 3 is a flow chart of an embodiment of a large model fine-tuning method according to the present disclosure;

[0018] Figure 4 is a schematic structural diagram of an embodiment of a data processing device according to the present disclosure;

[0019] Figure 5 is a structural schematic diagram of an embodiment of a large model fine-tuning device according to the present disclosure;

[0020] Figure 6 It is a block diagram of an electronic device used to implement the data processing method or large model fine-tuning method of the embodiment of the present disclosure. DETAILED DESCRIPTION

[0021] The following description of exemplary embodiments of the present disclosure is made in conjunction with the accompanying drawings, including various details of the embodiments of the present disclosure to facilitate understanding. These details should be considered as merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present disclosure. Similarly, for the sake of clarity and conciseness, descriptions of well-known functions and structures are omitted in the following description.

[0022] In the current technology landscape, most service platforms (such as live streaming platforms) employ the following mainstream solutions to address risky users: 1) Rule-based sensitive word censorship, which utilizes a sensitive word library combined with simple text censorship rules to capture sensitive words used in user comments. However, many sensitive word monitoring methods that rely on dictionary matching fail to capture highly variable attack language. For example, attackers may use complex internet jargon or expressions with implicit offensive meanings, which are easily missed by fixed sensitive word rule systems. Furthermore, word-banning strategies fail to address the overall semantics of a sentence.

[0023] 2) Manual review requires human intervention to screen and make decisions on user reports. However, manual review requires frame-by-frame analysis of large amounts of video content and comparing every text statement in the video clips. This is not only inefficient, but also increases the chance of misjudgment and delays processing time. In particular, it cannot meet the real-time review requirements in large-scale live broadcasts.

[0024] 3) Some systems also incorporate simple sentiment analysis models, such as positive / negative sentiment labeling systems, which use machine learning classification models to determine the offensiveness of a statement based on textual expression. However, these models are often targeted at short, limited-word reviews and tend to overlook context, historical comments, and even the user's overall background, making them prone to missed and false positives. However, most current basic sentiment analysis methods rely on static text sentiment classification and lack a deep understanding of the relationship between speech content and context. For users who engage in covert verbal attacks, single-sentence sentiment classification techniques struggle to capture the evolving trends of long-term speech patterns.

[0025] In response to the defects in traditional technologies, this paper proposes a data processing method that can automatically generate diagnostic scripts, improve user experience, Figure 1 A process 100 according to an embodiment of a data processing method of the present disclosure is shown. The data processing method includes the following steps:

[0026] Step 101: Acquire interactive message data and behavior operation data of a service object in a service platform.

[0027] In this embodiment, a service platform is a platform that provides services to both service recipients and service recipients, such as a live streaming platform. Service recipients are those who recommend products or services on the service platform, such as live streamers. Service recipients are those who receive services from service recipients on the service platform. The execution entity on which the data processing method runs can be the service platform, or it can be a monitoring system that monitors the service platform.

[0028] In this embodiment, the interactive message data of the service object is the information of information interaction between the service object and the service platform or the service object. Specifically, the interactive message data may include first interactive data of information interaction with the service object on goods or services, such as the service object consulting the service platform for product information of the goods that the service object is introducing; the interactive message data may also include second interactive data of information interaction with the service platform on goods or services, such as the service object reading product information on some display pages of the service platform.

[0029] In this embodiment, the behavioral operation data refers to the operation data of the service object in the service platform except for the interactive message data, such as the number of reports, likes, room joining frequency, attention behavior, etc.

[0030] Optionally, for service recipients who are dissatisfied with the service provided by the service object, the service platform provides a reporting mechanism that allows service recipients to report the service object's services to the service platform. Service recipients can report the poor service provided to the reporting service object by operating the reporting control on the service platform. When performing the reporting operation, the service recipient can provide feedback to the service platform including the reporting content data and the service object's service context data. Therefore, the behavioral operation data includes the reporting operation, and the interactive message data includes the reporting content data and service context data. The behavioral action of operating the reporting control is the reporting operation, the reporting content data is the text provided to the service platform by the service recipient after performing the reporting operation, and the service context data is the textual information exchanged between the service object and the service recipient.

[0031] When the service platform is a live broadcast platform, the execution entity on which the data processing method runs comprehensively analyzes the comprehensive behavior of the service object by analyzing the interaction messages between the service object and the live broadcast platform or the service object, combined with the behavioral operation data of the service object during the live broadcast.

[0032] The collection, storage, use, processing, transmission, provision, and disclosure of vehicle fault information in this disclosure are authorized and comply with relevant laws and regulations. Vehicle fault information and user-related information are obtained with the user's permission, and user-related data processing is conducted under confidentiality conditions.

[0033] like Figure 2As shown, the data collection module is used to collect the interactive message data and behavioral operation data of the service object, and input the interactive message data into the content understanding model and the behavioral operation data into the behavior prediction module respectively.

[0034] Step 102: Based on the interactive message data and the content understanding model, obtain the sentiment evaluation results and semantic sentiment labels.

[0035] In this embodiment, the emotion evaluation result is the result of evaluating the emotion of the service object, and the emotion evaluation result includes: the emotion expressed by the service object in the speech, such as positive, negative, neutral, etc.; language tendency information, such as whether it contains aggressive, sarcastic and other words; potential emotion intensification direction, which is inferred by analyzing the context to determine whether the user has emotional fluctuations, such as anger, loss of control or harassment.

[0036] In this embodiment, semantic emotion tags are used to label text corresponding to different text segments, different time periods, or different actions in the interactive message data with semantic emotions. Semantic emotion tags can be used to determine which speeches in the interactive message data belong to semantic categories (for example, support, neutrality, action, purpose, etc.) and which belong to emotional types (positive, negative, neutral). Semantic emotion tags include: positive semantic tags, negative semantic tags, emotional tags, action tags, and purpose tags. Positive semantic tags are used to identify text that praises and encourages; negative semantic tags are used to identify negative text; emotional tags are used to identify overly excited and irrational text; action tags are used to identify text with calls for action; and purpose tags are used to identify text with a purpose.

[0037] Optionally, for the reported content data and service context data in the interactive message data, the content understanding model can also classify the reported content data and service context data to obtain a reported content classification. The reported content classification is used to classify different credibility levels in the text and provide a basis for obtaining behavior evaluation results.

[0038] Step 103: Determine the behavior trajectory information of the service recipient based on the semantic emotion tag and the behavior operation data.

[0039] In this embodiment, the behavior trajectory information is information that combines a time series to reflect the operation mode of the served object. The behavior trajectory information can be used to predict whether the served object will develop into an attack target.

[0040] like Figure 2As shown in Figure 2, the behavior trajectory information of the service object is generated by the behavior prediction module. Since behavior cannot unilaterally reflect the attack behavior, the behavior prediction module does not determine the behavior trajectory information of the service object from the behavior operation data alone. Instead, it combines the semantic emotion tags to expand the various behavior operation data of the service object during the entire service process to generate the behavior trajectory information. Figure 2 In the process, the behavior prediction module receives the semantic emotion labels output by the content understanding model, generates behavior trajectory information, and inputs the behavior trajectory information, behavior operation data, and emotion evaluation results output by the content understanding model into the review module.

[0041] In this embodiment, the above-mentioned step 103 includes: using a time series neural network such as a long short-term memory network to store the behavioral operation data of the service object in the historical stage and the current stage through recursive memory, and combining the semantic emotion labels of each stage to generate the behavioral trajectory information of the service object represented by a sequence.

[0042] Step 104: Determine the behavior evaluation result of the service recipient based on the emotion evaluation result, the behavior operation data, and the behavior trajectory information.

[0043] In this embodiment, the behavior assessment results are based on the emotion assessment results, behavior operation data and behavior trajectory information, and comprehensively analyze the attack behavior of the service object from a time series perspective, and set further operations according to the service platform management strategy. The further operations include: automatic labeling, delayed processing, and direct intervention.

[0044] In this embodiment, the execution subject on which the data processing method runs includes an audit module, such as Figure 2 As shown, the review module is used to conduct a comprehensive analysis of the received emotion evaluation results, behavioral operation data, and behavioral trajectory information, and provide the behavioral evaluation results of the service recipients.

[0045] Optionally, the execution entity on which the data processing method runs also includes a processing feedback module for bad service objects. The behavior evaluation results of the service objects that are attack objects need to be sent to the feedback module, and the service objects are punished accordingly through the feedback module to achieve the purpose of purifying the service environment of the service platform.

[0046] The data processing method provided by the embodiment of the present disclosure first obtains the interactive message data and behavioral operation data of the service object in the service platform; secondly, based on the interactive message data and the content understanding large model, obtains the sentiment evaluation results and semantic sentiment labels; thirdly, based on the semantic sentiment labels and behavioral operation data, determines the behavioral trajectory information of the service object; finally, based on the sentiment evaluation results, behavioral operation data and behavioral trajectory information, determines the behavioral evaluation results of the service object. Thus, not only the sentiment analysis of the interactive message data of the service object is performed, but also the behavioral operation data of the service object on the service platform is combined with the global semantic and sentiment understanding of the context by using the large model, which can handle complex and ambiguous attack methods in the service platform, locate high-risk users, improve the accuracy of risk user detection, and reduce the damage of malicious attacks to the service platform ecology.

[0047] In some embodiments of the present disclosure, the data processing method further includes: sending processing prompt information in response to the behavior evaluation result indicating that the served object is an attack object.

[0048] In this embodiment, the processing prompt information is a prompt information that prompts the processing party to process the attack object of the service platform. The processing information is used to prompt the service platform to directly intervene in the service object, such as banning, suspending or reminding to exit the platform.

[0049] The data processing method provided by this optional implementation sends processing prompt information when the behavior evaluation result indicates that the service object is an attack target. The processing prompt information can effectively prompt the processing party to process the current attack target, thereby improving the security of the service platform environment.

[0050] Optionally, the above data processing method also includes: averaging the credibility levels under different actions in the behavior evaluation results to obtain an average credibility value; detecting whether the average credibility value is greater than a preset threshold; in response to detecting that the average credibility value is greater than the preset threshold, pushing processing suggestion information to the external feedback system or sending a system notification to the service object, wherein the processing suggestion information is used to suggest that the service platform ban the access rights of the service object so that the service object cannot enter the service platform; the system notification includes reminders and penalty results.

[0051] In some optional implementations of the present disclosure, the above-mentioned acquisition of the interactive message data and behavioral operation data of the served object in the service platform includes: obtaining the user identification of the served object in the service platform; based on the user identification, acquiring the interactive message data of the served object; based on the user identification, acquiring the behavioral operation data of the served object.

[0052] In this optional implementation, the interactive message data includes first interactive message data between the service object and the service platform, and second interactive message data between the service object and the service object. The data modalities of the interactive message data are in multiple forms, such as voice, text, and image.

[0053] In this optional implementation, the behavior operation data is information obtained after the service object operates the controls or components in the service platform. The behavior operation data includes: the operated object and the operation time, wherein the operation time includes the historical time and the current time.

[0054] In this optional implementation, the service platform assigns a unique user ID to each service provider and service recipient based on their respective identification codes. Each service recipient's interactive message data and behavioral operation data are associated with this unique user ID (such as a user ID or account number). Through this user ID, the service platform can accurately track the service recipient's historical comments and operational actions.

[0055] This optional implementation provides a method for obtaining the interactive message data and behavioral operation data of the service object in the service platform, obtaining the user identification of the service object in the service platform; based on the user identification, obtaining the interactive message data of the service object; based on the user identification, obtaining the behavioral operation data of the service object. By assigning a unique user identification to the service object, the interactive message data and behavioral operation data of the service object can be effectively obtained, thereby improving the reliability of obtaining the interactive message data and behavioral operation data.

[0056] Optionally, the above-mentioned acquisition of the interactive message data and behavioral operation data of the service object in the service platform includes: in response to the service platform being a live broadcast platform, obtaining the live broadcast room number of the live broadcast platform; for the user identifier of the service object in the live broadcast room number, obtaining the interactive message data and behavioral operation data corresponding to the user identifier.

[0057] In some optional implementations of the present disclosure, the above-mentioned acquisition of interactive message data and behavioral operation data of the service object in the service platform further includes: in response to the behavioral operation data including a reporting operation, identifying the reported content data in the interactive message data and the service context data of the service object.

[0058] In this optional implementation, the reporting operation is an operation in which the service object reports the service object. Generally, a reporting operation control is provided on the service platform, and the service object reports the service object as a dishonest user to the service platform by operating the reporting operation control.

[0059] In this optional implementation, the reported content data is the reported information conveyed by the service recipient to the service platform. By performing sentiment and semantic analysis on the reported content data, the attitude, sentiment and other information of the service recipient towards the service recipient can be determined.

[0060] The method for obtaining interactive message data and behavioral operation data provided by this optional implementation method, when the behavioral operation data includes a reporting operation, identifies the reported content data and service context data of the service object in the interactive message data, and can provide a reliable basis for subsequent analysis of the interactive message data and behavioral operation data.

[0061] In this embodiment, the above-mentioned emotion evaluation results and semantic emotion labels based on interactive message data and content understanding big model include: inputting reported content data and service context data into the content understanding big model to obtain the reported content classification of the reported content data and the semantic emotion labels of the service context data.

[0062] In some optional implementations of the present disclosure, the above-mentioned obtaining of sentiment evaluation results and semantic sentiment labels based on interactive message data and the content understanding big model includes: performing text preprocessing on the interactive message data to obtain preprocessed data; inputting the preprocessed data into the content understanding big model to obtain sentiment evaluation results and semantic sentiment labels output by the content understanding big model.

[0063] In this optional implementation, text preprocessing of interactive message data includes: cleaning and preprocessing the interactive message data, which includes removing redundant characters, filtering stop words, correcting spellings, and processing punctuation marks. Text preprocessing can ensure that the preprocessed data has a structure that can be processed efficiently and avoid noise interference.

[0064] In this optional implementation, the content understanding model can be a large language model based on the Transformer architecture (such as BERT or the GPT series of models). The preprocessed data segments are sequentially processed within the model's multi-layered deep neural network. The model includes a multi-layered attention mechanism, which allows it to understand the importance of each word in the preprocessed data within the context of the sentence and its semantic connections with other words. Furthermore, the model generates sentiment evaluation results and semantic sentiment labels.

[0065] In this optional implementation, the semantic emotion tag is a tag that identifies the semantics or emotion of each segment of data in the interactive message data. The semantic emotion tag can provide a reliable analysis basis for the analysis of the behavioral operation data corresponding to the interactive message data.

[0066] This optional implementation provides a method for obtaining sentiment assessment results and semantic sentiment labels. This method preprocesses interactive message data to obtain preprocessed data. This preprocessed data is then fed into a content understanding model to generate sentiment assessment results and semantic sentiment labels. This model can then perform a holistic semantic and sentiment understanding of the interactive message data, labeling each segment within the interactive message data with a corresponding semantic sentiment label, thus improving the comprehensiveness of the interactive message data analysis.

[0067] Optionally, the above-mentioned sentiment evaluation results and semantic sentiment labels based on the interactive message data and the content understanding big model include: performing text preprocessing on the interactive message data to obtain preprocessed data; inputting the preprocessed data into the content understanding big model to obtain the sentiment evaluation results output by the content understanding big model; inputting the preprocessed data into the natural language processing model to obtain the semantic sentiment labels output by the natural language processing model.

[0068] Optionally, before feeding the preprocessed data into the content understanding model, it can be segmented. Segmentation involves breaking the preprocessed data into words or phrases, which are then converted into specific vector representations. Text embedding techniques can be used to convert text into multidimensional vectors, where the relationships between individual words and phrases can be expressed through relative positions and similarities between word vectors.

[0069] In some optional implementations of the present disclosure, the above-mentioned determination of the behavioral trajectory information of the served object based on semantic emotion tags and behavioral operation data includes: determining the operation frequency of the target action of the served object based on the behavioral operation data; generating the behavioral trajectory information of the served object based on the semantic emotion tags, operation frequency and behavioral operation data.

[0070] In this optional implementation, the target action is an action taken by the service recipient on the service platform due to dissatisfaction with the service recipient's service, such as entering the service platform or clicking a report button. When the frequency of the target action exceeds a certain threshold, the service recipient is determined to be suspected of attacking the service recipient.

[0071] In this optional implementation, the above-mentioned generation of behavioral trajectory information of the service object based on semantic emotion labels, operation frequency and behavioral operation data includes: detecting whether the operation frequency exceeds the frequency threshold; in response to detecting that the operation frequency exceeds the frequency threshold, detecting whether the semantic emotion label corresponding to the target action is a negative semantic label, if it is a message semantic label, concatenating the actions of the service object in the behavioral operation data and the semantic emotion labels corresponding to each action to obtain the behavioral trajectory information of the service object.

[0072] The method for determining behavior trajectory information provided by this optional implementation method determines the operation frequency of the target action of the service object based on the behavior operation data; generates the behavior trajectory information of the service object based on the semantic emotion label, the operation frequency and the behavior operation data. Therefore, based on the behavior operation data, the operation frequency is determined, which can effectively determine the repetition of the operation behavior of the service object on the service platform; based on the semantic emotion label and the operation frequency, the emotional color of the service object frequently operated on the service platform can be obtained, which provides a reliable basis for the aggressive behavior of the service object and improves the accuracy of the behavior trajectory information.

[0073] Optionally, the above-mentioned determination of the behavior trajectory information of the service object based on semantic emotion tags and behavior operation data includes: based on the data segment in the interactive message data corresponding to the semantic emotion tags; determining the behavior operation data corresponding to the data segment, and marking the behavior operation data with semantic emotion tags; combining different time periods in the behavior operation data with semantic emotion tags to obtain the behavior trajectory information of the service object.

[0074] In some optional implementations of the present disclosure, the above-mentioned behavioral operation data includes: current operation data and historical operation data. Based on the behavioral operation data, determining the operation frequency of the target action of the service object includes: inputting the current operation data and the historical operation data into a time series model to obtain the behavior change results output by the time series model; based on the behavior change results, determining the operation frequency of the target action of the service object.

[0075] In this optional implementation, the current operation data and the historical operation data can be obtained by dividing the behavior operation data at a specific time point. The current operation data and the historical operation data are both operation data corresponding to the same service object.

[0076] In this optional implementation, the time series model is used to analyze the time series changes of the current operation data and the historical operation data.

[0077] This optional implementation provides a method for determining the operating frequency of the target action of the service object, which inputs current operating data and historical operating data into a time series model to obtain a behavior change result output by the time series model; based on the behavior change result, the operating frequency of the target action of the service object is determined, thereby effectively reflecting the operating frequency of the target action through time series analysis of the historical operating data and the current operating data, thereby improving the accuracy and reliability of the obtained operating frequency.

[0078] In some optional implementations of the present disclosure, the above-mentioned generation of behavioral trajectory information of the served object based on semantic emotion labels, operation frequencies and behavioral operation data includes: detecting whether the semantic emotion labels match the operation frequencies; in response to detecting that the semantic emotion labels match the operation frequencies, determining the contextual behavior frequency based on the operation frequencies; generating the behavioral trajectory information of the served object based on the contextual behavior frequency and the behavioral operation data; in response to detecting that the semantic emotion labels do not match the operation frequencies, filtering information reflecting attack phenomena in the semantic emotion labels and the operation frequencies; in response to filtering out information reflecting attack phenomena, generating the behavioral trajectory information of the served object based on the behavioral operation data and the filtering identifier, wherein the filtering identifier is used to identify the served object as suspected of being an attack object.

[0079] In this optional implementation, the contextual behavior frequency is the frequency of behavioral operations of the service object at different interaction stages on the service platform. For example, five minutes ago, the service object frequently entered and exited the service platform, and five minutes later, the service object was fixed in the service platform to interact with the service object. At this time, the contextual behavior frequency includes two, one is the entry and exit frequency five minutes ago, and the other is the fixed frequency five minutes later.

[0080] In this optional implementation, the behavior operation data may be divided into multiple data segments according to the context behavior frequency, and the summary information corresponding to each data segment is a track point in the behavior track information.

[0081] In this optional implementation, the information reflecting the attack phenomenon refers to the frequency at which the semantic emotion label is an offensive label or the operation frequency hits the offensive action; the above-mentioned screening identifier includes: a first identifier and a second identifier, wherein the first identifier is used to identify that the service object is suspected to be the attack object through the behavioral representation of the service object, and the second identifier is used to identify that the service object is suspected to be the attack object through the language representation of the service object. The above-mentioned response to detecting that the semantic emotion label does not match the operation frequency, screening the information reflecting the attack phenomenon in the semantic emotion label and the operation frequency, and in response to screening the information reflecting the attack phenomenon, generating the behavior trajectory information of the service object based on the behavioral operation data and the screening identifier includes: in response to detecting that the semantic emotion label does not match the operation frequency, screening the information reflecting the attack phenomenon in the semantic emotion label and the operation frequency, and in response to screening the operation frequency reflecting the attack phenomenon, generating the behavior trajectory information of the service object based on the behavioral operation data and the first identifier.

[0082] In this optional implementation, the above-mentioned response to detecting that the semantic emotion label does not match the operating frequency, filtering the information reflecting the attack phenomenon in the semantic emotion label and the operating frequency, and in response to filtering out the information reflecting the attack phenomenon, generating the behavior trajectory information of the served object based on the behavioral operation data and the filtering identifier includes: in response to detecting that the semantic emotion label does not match the operating frequency, filtering the information reflecting the attack phenomenon in the semantic emotion label and the operating frequency, and in response to filtering out the semantic emotion label reflecting the attack phenomenon, generating the behavior trajectory information of the served object based on the behavioral operation data and the second identifier.

[0083] The method for generating behavioral trajectory information of the service object provided by this optional implementation method determines the contextual behavior frequency based on the operation frequency when the semantic emotion label matches the operation frequency; generates the behavioral trajectory information of the service object based on the contextual behavior frequency and behavioral operation data. The generation of behavioral trajectory information through the contextual behavior frequency provides a reliable technical means for the effective division of behavioral trajectory information, thereby improving the reliability of obtaining behavioral trajectory information.

[0084] In some optional implementations of the present disclosure, determining the behavioral evaluation result of the served object based on the emotion evaluation result, behavioral operation data, and behavioral trajectory information includes: detecting whether the emotion evaluation result matches the behavioral trajectory information; in response to the emotion evaluation result matching the behavioral trajectory information, generating a multidimensional behavior evaluation table based on the emotion evaluation result, behavioral operation data, and behavioral trajectory information; generating a behavioral evaluation result of whether the served object is an attacking object or a non-attacking object based on the multidimensional behavior evaluation table; in response to the emotion evaluation result not matching the behavioral trajectory information, generating an emotion development curve based on the emotion evaluation result; generating a behavior change curve based on the behavioral trajectory information and the behavioral operation data; in response to the emotion development curve and the behavior change curve not exceeding their respective curve limits, generating a behavior evaluation result of the served object as a non-attacking object, wherein the curve limits are the emotion and behavior limit ranges set for the served object.

[0085] In this optional implementation method, the emotion evaluation result reflects the emotion or feeling change trend information of the service object from the speech of the service object, and the behavior trajectory information reflects the emotion or feeling change trend information of the service object from the behavior of the service object. When the trend of the emotion evaluation result and the behavior trajectory information are the same, it is determined that the service object has launched an attack on the service object in both words and actions, which is an attack on the service object.

[0086] In this optional implementation, the multidimensional behavior evaluation table is a table that reflects the credibility of the service recipient from multiple dimensions. The emotional evaluation results and behavioral trajectory information in different time periods have corresponding credibility. For service recipients with attack tendencies, the credibility reflected in their multidimensional behavior evaluation table gradually decreases. For service recipients who reach a certain credibility trend, they can be directly confirmed as attack targets. Specifically, the multidimensional behavior evaluation table can include the service recipient's historical operation records, reporting logic, user activity data, etc. Among them, the information in the multidimensional behavior evaluation table can be obtained by analyzing the behavioral operation data of the service recipient.

[0087] In this optional implementation, the above-mentioned generation of a behavior evaluation result of whether the service object is an attack object or a non-attack object based on the multidimensional behavior evaluation table includes: establishing different grading standards for actions from different sources and categories in the multidimensional behavior evaluation table, and when the standards of the attack object are met in the multidimensional behavior evaluation table, determining that the service object is an attack object; otherwise, determining that the service object is a non-attack object.

[0088] This optional implementation provides a method for determining a behavioral assessment result for a service recipient. The method detects whether the emotion assessment result matches the behavioral trajectory information. In response to a match between the emotion assessment result and the behavioral trajectory information, a multidimensional behavior assessment table is generated based on the emotion assessment result, the behavioral operation data, and the behavioral trajectory information. Based on the multidimensional behavior assessment table, a behavior assessment result is generated indicating whether the service recipient is an attack target or a non-attack target. Thus, the reliability of the behavior assessment result is improved by reflecting whether the service recipient is an attack target or a non-attack target through the multidimensional behavior assessment table.

[0089] In some optional implementations of the present disclosure, the emotion assessment result includes: reporting content classification, the above-mentioned response to the emotion assessment result matching the behavior trajectory information, generating a multidimensional behavior assessment table based on the emotion assessment result, behavior operation data and behavior trajectory information includes: responding to the emotion assessment result matching the behavior trajectory information, based on the behavior trajectory information, determining historical operation actions from the behavior operation data; generating a multidimensional behavior assessment table based on historical operation actions and reporting content classification.

[0090] In this optional implementation, the reported content classification is the data obtained after classifying the reported content data of the service object. The interactive message data contains an identifier representing the reported content data. The content understanding model performs sentiment classification on the reported content data to determine the reported content classification.

[0091] In this optional implementation, the above-mentioned generation of a multidimensional behavior evaluation table based on historical operation actions and reported content classification includes: aggregating the reported content classification and historical operation actions belonging to the same type together to obtain action sets under different clusters; grading the action sets under different clusters according to corresponding grading standards to obtain a multidimensional behavior evaluation table.

[0092] The method for generating a multidimensional behavior evaluation table provided by this optional implementation method responds to the matching of the emotion evaluation results with the behavior trajectory information, determines the historical operation actions from the behavior operation data based on the behavior trajectory information; generates a multidimensional behavior evaluation table based on the historical operation actions and the classification of the reported content. By generating the multidimensional behavior evaluation table through the reported content, the evaluation of the behavior of the service object can be refined, thereby improving the reliability of the multidimensional behavior evaluation table.

[0093] Figure 3 A process 300 according to an embodiment of a large model fine-tuning method of the present disclosure is shown. The large model fine-tuning method includes the following steps:

[0094] Step 301: Obtain a training sample set and a pre-trained large model.

[0095] In this embodiment, the training sample set includes at least one training sample, and the training sample includes: sample message data and multi-task annotation data corresponding to the sample message data.

[0096] In this embodiment, the sample message data is data collected within a specific field, such as data in a live broadcast platform scenario. The sample message data includes: new user speech sample data and new popular speech styles in different time periods. The multi-task annotation data includes: content classification labels and contextual semantic labels. Among them, the content classification labels include reported content classification labels. The content classification labels can provide an effective basis for fine-tuning the sentiment evaluation results generated by the pre-trained large model, and the contextual semantic labels can provide an effective basis for fine-tuning the semantic sentiment labels generated by the pre-trained large model.

[0097] In this embodiment, the pre-trained large model is obtained by pre-training an initial large-scale natural language processing model (for example, a Transformer-based model). For different service platforms, the pre-trained large model can be pre-trained in combination with the historical data of the service platform. By combining this historical data with the initial large-scale natural language processing model, a pre-trained large model with language processing capabilities for live broadcast scenarios is initially obtained.

[0098] In this embodiment, for live streaming platforms, historical data includes speech content data and user behavior data. Speech content data refers to the chat information of users in the live streaming room, which generally includes a large number of positive interactions and negative speeches. The goal is to enable the large model to grasp the interaction patterns of users on the live streaming platform during the pre-training phase. User behavior data includes the user's behavior trajectory during the live streaming process, such as operation data (likes, shares, reports, number of times entering and exiting the live streaming room, etc.).

[0099] In the technical solution disclosed herein, the collection, storage, use, processing, transmission, provision and disclosure of the training sample sets and pre-trained large models involved are carried out after authorization and in compliance with relevant laws and regulations.

[0100] Step 302: Select training samples from the training sample set.

[0101] In this embodiment, the execution entity on which the large model fine-tuning method runs can randomly select training samples from the training samples, or select fixed-form training samples from the training samples. The selection method and number of training samples from the training sample set are not restricted in this disclosure, and the number of iterative training of the pre-trained large model is not restricted.

[0102] Step 303: Input the selected training samples into the large model to obtain the sentiment evaluation results and semantic sentiment labels output by the large model.

[0103] In this embodiment, the pre-trained large model updates itself from continuously generated new data, enhancing its understanding of complex cases through new labels and new data (such as administrators' annotated data). For example, when new types of verbal attacks or combined spam are captured, the large model needs to incorporate these new types of content to adapt to the new challenges.

[0104] Step 304: Based on the sentiment evaluation results, semantic sentiment labels, and selected training samples, the large model is trained to obtain a trained content understanding large model.

[0105] In this embodiment, the above-mentioned training of the large model refers to continuous fine-tuning of the large model. The fine-tuning method is a conventional method and will not be repeated here. As the training samples continue to change, fine-tuning is not a one-time process, but a continuous process. This part achieves dynamic fine-tuning through incremental learning and self-supervised learning.

[0106] In this embodiment, the trained content understanding model is continuously trained based on a large amount of historical chat data, various types of reported events, and known cases of illegal content in the community. As the amount of data samples increases, the model's label generation ability will become more accurate.

[0107] In this embodiment, the trained content understanding model supports adaptive dynamic learning. For example, an administrator can add feedback on misclassification situations. After the correction process, the model will better adjust its own predictions in future scenarios.

[0108] In a live streaming scenario, when a user makes a false report to the live streaming platform (such as reporting "illegal content" to the anchor), the content understanding model will first extract the user's speech and operation behavior through context analysis. After fine-tuning the language, the model combines the context and finds that the chat before and after the speech did not show actual concern for the "illegal content", but rather showed a derogatory emotional attack. Therefore, the model predicts that this is a negative emotional evaluation result. Secondly, the content understanding model further considers the user's past behavior trajectory and the historical operation data of the live streaming room through a fine-tuned behavior prediction mechanism, and recognizes that the user has a history of frequent malicious reports. Therefore, the report is labeled with the semantic emotion label of "malicious and false."

[0109] The large model fine-tuning method provided in this embodiment trains the pre-trained large model based on the training samples in the training sample set, so that the large model can fully learn the multi-task annotation data in the training samples, output sentiment evaluation results and semantic sentiment labels, and provide a reliable implementation method for obtaining the content understanding large model, thereby improving the reliability of the content understanding large model.

[0110] Further references Figure 4 As an implementation of the methods shown in the above figures, the present disclosure provides an embodiment of a data processing device, which is similar to Figure 1 Corresponding to the method embodiment shown, the device can be specifically implemented in various electronic devices.

[0111] like Figure 4 As shown, the data processing device 400 provided in this embodiment includes: a data acquisition unit 401, an obtaining unit 402, a trajectory determination unit 403, and a result determination unit 404. Among them, the above-mentioned data acquisition unit 401 can be configured to obtain the interactive message data and behavioral operation data of the service object in the service platform. The above-mentioned obtaining unit 402 can be configured to obtain the emotion evaluation result and the semantic emotion label based on the interactive message data and the content understanding model. The above-mentioned trajectory determination unit 403 can be configured to determine the behavior trajectory information of the service object based on the semantic emotion label and the behavior operation data. The above-mentioned result determination unit 404 can be configured to determine the behavior evaluation result of the service object based on the emotion evaluation result, the behavior operation data and the behavior trajectory information.

[0112] In this embodiment, the specific processing of the data acquisition unit 401, the obtaining unit 402, the trajectory determination unit 403, and the result determination unit 404 and the technical effects thereof can be referred to in the respective Figure 1 The relevant descriptions of step 101, step 102, step 103, and step 104 in the corresponding embodiment are not repeated here.

[0113] In some optional implementations of this embodiment, the apparatus 400 further includes a sending unit (not shown in the figure), which may be configured to send processing prompt information in response to the behavior evaluation result indicating that the served object is an attacking object.

[0114] In some optional implementations of this embodiment, the above-mentioned data acquisition unit 401 is configured to: obtain the user identification of the service object in the service platform; based on the user identification, obtain the interactive message data of the service object; based on the user identification, obtain the behavioral operation data of the service object.

[0115] In some optional implementations of this embodiment, the acquisition unit 401 is further configured to: in response to the behavior operation data including a reporting operation, identify the reporting content data and service context data of the service object in the interaction message data.

[0116] In some optional implementations of this embodiment, the above-mentioned obtaining unit 402 is configured to: perform text preprocessing on the interactive message data to obtain preprocessed data; input the preprocessed data into the content understanding big model to obtain the sentiment evaluation results and semantic sentiment labels output by the content understanding big model.

[0117] In some optional implementations of this embodiment, the above-mentioned trajectory determination unit 403 is configured to: determine the operation frequency of the target action of the service object based on the behavioral operation data; and generate behavioral trajectory information of the service object based on the semantic emotion label, operation frequency and behavioral operation data.

[0118] In some optional implementations of this embodiment, the above-mentioned behavioral operation data includes: current operation data and historical operation data, and the above-mentioned trajectory determination unit 403 is further configured to: input the current operation data and historical operation data into a time series model to obtain the behavior change results output by the time series model; based on the behavior change results, determine the operation frequency of the target action of the service object.

[0119] In some optional implementations of this embodiment, the above-mentioned trajectory determination unit 403 is further configured to: detect whether the semantic emotion label matches the operation frequency; in response to detecting that the semantic emotion label matches the operation frequency, determine the context behavior frequency based on the operation frequency; generate behavior trajectory information of the served object based on the context behavior frequency and the behavior operation data; in response to detecting that the semantic emotion label does not match the operation frequency, filter the information reflecting the attack phenomenon in the semantic emotion label and the operation frequency; in response to filtering out the information reflecting the attack phenomenon, generate behavior trajectory information of the served object based on the behavior operation data and the filtering identifier, and the filtering identifier is used to identify the served object as suspected of being an attack object.

[0120] In some optional implementations of this embodiment, the above-mentioned result determination unit 404 is configured to: detect whether the emotion evaluation result matches the behavior trajectory information; in response to the emotion evaluation result matching the behavior trajectory information, generate a multidimensional behavior evaluation table based on the emotion evaluation result, behavior operation data and behavior trajectory information; based on the multidimensional behavior evaluation table, generate a behavior evaluation result of whether the service object is an attack object or a non-attack object; in response to the emotion evaluation result not matching the behavior trajectory information, generate an emotion development curve based on the emotion evaluation result; generate a behavior change curve based on the behavior trajectory information and the behavior operation data; in response to both the emotion development curve and the behavior change curve not exceeding their respective curve limits, generate a behavior evaluation result of the service object being a non-attack object, wherein the curve limits are the emotion and behavior limit ranges set for the service object.

[0121] In some optional implementations of the present disclosure, the above-mentioned emotion evaluation results include: reported content classification, and the above-mentioned result determination unit 404 is configured to: in response to the emotion evaluation results matching the behavior trajectory information, determine the historical operation actions from the behavior operation data based on the behavior trajectory information; generate a multidimensional behavior evaluation table based on the historical operation actions and the reported content classification.

[0122] The data processing device provided by the embodiment of the present disclosure, first, the data acquisition unit 401 acquires the interactive message data and behavioral operation data of the service object in the service platform; secondly, the acquisition unit 402 obtains the sentiment evaluation result and the semantic sentiment label based on the interactive message data and the content understanding large model; thirdly, the trajectory determination unit 403 determines the behavioral trajectory information of the service object based on the semantic sentiment label and the behavioral operation data; finally, the result determination unit 404 determines the behavioral evaluation result of the service object based on the sentiment evaluation result, the behavioral operation data and the behavioral trajectory information. Thus, not only the sentiment analysis is performed on the interactive message data of the service object, but also the behavioral operation data of the service object on the service platform is combined with the global semantic and sentiment understanding of the context by using the large model, which can handle complex and ambiguous attack methods in the service platform, locate high-risk users, improve the accuracy of risk user detection, and reduce the damage of malicious attacks to the service platform ecology.

[0123] In the technical solutions disclosed herein, the collection, storage, use, processing, transmission, provision and disclosure of user personal information involved comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0124] Further references Figure 5 As an implementation of the methods shown in the above figures, the present disclosure provides an embodiment of a large model fine-tuning device, which is similar to Figure 1 Corresponding to the method embodiment shown, the device can be specifically implemented in various electronic devices.

[0125] like Figure 5 As shown, the large model fine-tuning device 500 provided in this embodiment includes: an information acquisition unit 501, a selection unit 502, an input unit 503, and a training unit 504. The information acquisition unit 501 can be configured to obtain a training sample set and a pre-trained large model, and the training sample set includes at least one training sample, and the training sample includes: sample message data and multi-task annotation data corresponding to the sample message data. The selection unit 502 can be configured to select training samples from the training sample set. The input unit 503 can be configured to input the selected training samples into the large model to obtain the sentiment evaluation results and semantic sentiment labels output by the large model. The training unit 504 can be configured to train the large model based on the sentiment evaluation results, semantic sentiment labels and selected training samples to obtain a trained content understanding large model.

[0126] In this embodiment, the specific processing of the information acquisition unit 501, the selection unit 502, the input unit 503, and the training unit 504 and the technical effects thereof can be referred to in detail. Figure 3The relevant descriptions of step 301, step 302, step 303, and step 304 in the corresponding embodiment are not repeated here.

[0127] According to an embodiment of the present disclosure, the present disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0128] Figure 6 A schematic block diagram of an example electronic device 600 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are provided as examples only and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0129] like Figure 6 As shown, the device 600 includes a computing unit 601, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 602 or a computer program loaded from a storage unit 608 into a random access memory (RAM) 603. Various programs and data required for the operation of the device 600 can also be stored in the RAM 603. The computing unit 601, the ROM 602, and the RAM 603 are connected to each other via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0130] Various components in device 600 are connected to I / O interface 605, including an input unit 606, such as a keyboard, mouse, etc.; an output unit 607, such as various types of displays, speakers, etc.; a storage unit 608, such as a magnetic disk, optical disk, etc.; and a communication unit 609, such as a network card, modem, wireless communication transceiver, etc. The communication unit 609 allows device 600 to exchange information / data with other devices via a computer network such as the Internet and / or various telecommunication networks.

[0131] The computing unit 601 can be a variety of general-purpose and / or specialized processing components with processing and computing capabilities. Some examples of the computing unit 601 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various computing units that run machine learning model algorithms, a digital signal processor (DSP), and any appropriate processor, controller, microcontroller, etc. The computing unit 601 performs the various methods and processes described above, such as the data processing method or the large model fine-tuning method. For example, in some embodiments, the data processing method or the large model fine-tuning method can be implemented as a computer software program that is tangibly contained in a machine-readable medium, such as the storage unit 608. In some embodiments, part or all of the computer program can be loaded and / or installed on the device 600 via the ROM 602 and / or the communication unit 609. When the computer program is loaded into the RAM 603 and executed by the computing unit 601, one or more steps of the data processing method or the large model fine-tuning method described above can be performed. Alternatively, in other embodiments, the computing unit 601 may be configured to execute the data processing method or the large model fine-tuning method in any other appropriate manner (for example, by means of firmware).

[0132] Various embodiments of the systems and techniques described herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), system-on-chip systems (SOCs), programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments can include being implemented in one or more computer programs that are executable and / or interpreted on a programmable system that includes at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0133] The program code for implementing the methods of the present disclosure may be written in any combination of one or more programming languages. Such program code may be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device or a large-scale model fine-tuning device, such that when the program code is executed by the processor or controller, the functions / operations specified in the flowcharts and / or block diagrams are implemented. The program code may be executed entirely on the machine, partially on the machine, as a stand-alone software package, partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0134] In the context of the present disclosure, a machine-readable medium can be a tangible medium that can contain or store a program for use by or in conjunction with an instruction execution system, device or equipment. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium can include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0135] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0136] The systems and techniques described herein can be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an information server), or a computing system that includes front-end components (e.g., a user computer having a graphical user interface or a web browser through which a user can interact with embodiments of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network (LAN), a wide area network (WAN), and the Internet.

[0137] A computer system may include a client and a server. The client and server are generally remote from each other and typically interact through a communication network. The client-server relationship arises through computer programs running on the respective computers and having a client-server relationship with each other. The server may be a cloud server, a server in a distributed system, or a server integrated with a blockchain.

[0138] It should be understood that the various forms of the processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this disclosure can be performed in parallel, sequentially, or in a different order, as long as the desired results of the technical solutions disclosed in this disclosure can be achieved. This is not limited herein.

[0139] The above specific embodiments do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and substitutions may be made based on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure shall be included within the scope of protection of this disclosure.

Claims

1. A data processing method, comprising: Obtain interactive message data and behavioral operation data of service objects in the service platform; Based on the interactive message data and the content understanding model, obtaining sentiment evaluation results and semantic sentiment labels; Determining the behavior trajectory information of the served object based on the semantic emotion tag and the behavior operation data; Determining a behavior evaluation result of the service recipient based on the emotion evaluation result, the behavior operation data, and the behavior trajectory information; The determining of the behavior trajectory information of the served object based on the semantic emotion tag and the behavior operation data includes: Determining the operation frequency of the target action of the served object based on the behavioral operation data; Generate behavior trajectory information of the served object based on the semantic emotion tag, the operation frequency, and the behavior operation data, including: Detecting whether the semantic emotion label matches the operation frequency; In response to detecting that the semantic emotion tag matches the operation frequency, determining a context behavior frequency based on the operation frequency; and generating behavior trajectory information of the served object based on the context behavior frequency and the behavior operation data; In response to detecting that the semantic emotion tag does not match the operating frequency, the information reflecting the attack phenomenon in the semantic emotion tag and the operating frequency is filtered. In response to filtering out the information reflecting the attack phenomenon, the behavior trajectory information of the served object is generated based on the behavioral operation data and the filtering identifier, and the filtering identifier is used to identify that the served object is suspected to be an attack object.

2. The method according to claim 1, further comprising: In response to the behavior evaluation result indicating that the served object is an attacking object, processing prompt information is sent.

3. The method according to claim 1 or 2, wherein: The acquisition of interactive message data and behavioral operation data of the service object in the service platform includes: Obtain the user ID of the service object in the service platform; Based on the user identifier, acquiring the interactive message data of the served object; Based on the user identifier, behavioral operation data of the served object is obtained.

4. The method according to claim 3, wherein: The acquisition of interactive message data and behavioral operation data of the service object in the service platform also includes: In response to the behavior operation data including a reporting operation, reporting content data and service context data of a service object in the interaction message data are identified.

5. The method according to claim 1 or 2, wherein: The obtaining of sentiment evaluation results and semantic sentiment labels based on the interactive message data and the content understanding large model includes: Performing text preprocessing on the interactive message data to obtain preprocessed data; The preprocessed data is input into the content understanding model to obtain the sentiment evaluation results and semantic sentiment labels output by the content understanding model.

6. The method according to claim 1, wherein The behavior operation data includes: current operation data and historical operation data. The determining the operation frequency of the target action of the served object based on the behavior operation data includes: Inputting the current operation data and the historical operation data into a time series model to obtain a behavior change result output by the time series model; Based on the behavior change result, the operation frequency of the target action of the served object is determined.

7. The method according to claim 1 or 2, wherein: The determining of the behavior evaluation result of the service object based on the emotion evaluation result, the behavior operation data, and the behavior trajectory information includes: Detecting whether the emotion assessment result matches the behavior trajectory information; In response to the emotion evaluation result matching the behavior trajectory information, generating a multidimensional behavior evaluation table based on the emotion evaluation result, the behavior operation data, and the behavior trajectory information; generating a behavior evaluation result of whether the served object is an attacking object or a non-attacking object based on the multidimensional behavior evaluation table; In response to the emotion assessment result not matching the behavior trajectory information, an emotion development curve is generated based on the emotion assessment result; a behavior change curve is generated based on the behavior trajectory information and the behavior operation data; in response to the emotion development curve and the behavior change curve not exceeding their respective curve limits, a behavior assessment result is generated indicating that the served object is a non-aggressive object, wherein the curve limits are the emotion and behavior limit ranges set for the served object.

8. The method according to claim 7, wherein: The emotion evaluation result includes: reporting content classification, the response matching the emotion evaluation result with the behavior trajectory information, generating a multidimensional behavior evaluation table based on the emotion evaluation result, the behavior operation data and the behavior trajectory information includes: In response to the emotion evaluation result matching the behavior trajectory information, determining a historical operation action from the behavior operation data based on the behavior trajectory information; Based on the historical operation actions and the reported content classification, a multi-dimensional behavior evaluation table is generated.

9. A large model fine-tuning method, the method comprising: Obtaining a training sample set and a pre-trained large model, wherein the training sample set includes at least one training sample, and the training sample includes: sample message data and multi-task annotation data corresponding to the sample message data; Selecting training samples from the training sample set; Input the selected training samples into the large model to obtain the sentiment evaluation results and semantic sentiment labels output by the large model; Based on the emotion evaluation result, the semantic emotion label and the selected training samples, the large model is trained to obtain a trained content understanding large model; The content understanding model is used in the method described in claim 1 to obtain sentiment evaluation results and semantic sentiment labels.

10. A data processing device, comprising: A data acquisition unit is configured to acquire interactive message data and behavioral operation data of a service object in the service platform; an obtaining unit configured to obtain a sentiment evaluation result and a semantic sentiment label based on the interactive message data and the content understanding large model; a trajectory determination unit configured to determine the behavior trajectory information of the served object based on the semantic emotion tag and the behavior operation data; A result determination unit is configured to determine a behavior evaluation result of the service object based on the emotion evaluation result, the behavior operation data and the behavior trajectory information; The trajectory determination unit is configured to: determine the operation frequency of the target action of the served object based on the behavioral operation data; generating behavior trajectory information of the served object based on the semantic emotion tag, the operation frequency, and the behavior operation data; The trajectory determination unit is further configured to: detect whether the semantic emotion tag matches the operation frequency; in response to detecting that the semantic emotion tag matches the operation frequency, determine the context behavior frequency based on the operation frequency; generate the behavior trajectory information of the served object based on the context behavior frequency and the behavior operation data; In response to detecting that the semantic emotion tag does not match the operating frequency, the information reflecting the attack phenomenon in the semantic emotion tag and the operating frequency is filtered. In response to filtering out the information reflecting the attack phenomenon, the behavior trajectory information of the served object is generated based on the behavioral operation data and the filtering identifier, and the filtering identifier is used to identify that the served object is suspected to be an attack object.

11. The apparatus according to claim 10, further comprising: The sending unit is configured to send processing prompt information in response to the behavior evaluation result indicating that the served object is an attack object.

12. The device according to claim 10 or 11, wherein The data acquisition unit is configured to: acquire a user identification of a service object in a service platform; acquire interactive message data of the service object based on the user identification; and acquire behavioral operation data of the service object based on the user identification.

13. The device according to claim 12, wherein The acquiring unit is further configured to: in response to the behavior operation data including a reporting operation, identify the reporting content data in the interaction message data and the service context data of the service object.

14. The device according to claim 10 or 11, wherein The obtaining unit is configured to: perform text preprocessing on the interactive message data to obtain preprocessed data; input the preprocessed data into the content understanding model to obtain the sentiment evaluation result and semantic sentiment label output by the content understanding model.

15. The device according to claim 10, wherein The behavioral operation data includes: current operation data and historical operation data. The trajectory determination unit is further configured to: input the current operation data and the historical operation data into a time series model to obtain a behavior change result output by the time series model; and determine the operation frequency of the target action of the service object based on the behavior change result.

16. The device according to claim 10 or 11, wherein The result determination unit is configured to: detect whether the emotion evaluation result matches the behavior trajectory information; in response to the emotion evaluation result matching the behavior trajectory information, generate a multidimensional behavior evaluation table based on the emotion evaluation result, the behavior operation data, and the behavior trajectory information; Based on the multi-dimensional behavior evaluation table, generating a behavior evaluation result of whether the served object is an attack object or a non-attack object; In response to the emotion assessment result not matching the behavior trajectory information, generating an emotion development curve based on the emotion assessment result; Based on the behavior trajectory information and the behavior operation data, a behavior change curve is generated; in response to the fact that neither the emotion development curve nor the behavior change curve exceeds their respective curve limits, a behavior assessment result is generated that the served object is a non-aggressive object, wherein the curve limits are the emotion and behavior limit ranges set for the served object.

17. The device according to claim 16, wherein The emotion assessment result includes: a reported content classification, and the result determination unit is configured to: in response to the emotion assessment result matching the behavior trajectory information, determine historical operation actions from the behavior operation data based on the behavior trajectory information; and generate a multidimensional behavior assessment table based on the historical operation actions and the reported content classification.

18. A large model fine-tuning device, comprising: An information acquisition unit is configured to acquire a training sample set and a pre-trained large model, wherein the training sample set includes at least one training sample, and the training sample includes: sample message data and multi-task annotation data corresponding to the sample message data; A selection unit, configured to select a training sample from the training sample set; An input unit is configured to input the selected training samples into the large model to obtain the sentiment evaluation results and semantic sentiment labels output by the large model; A training unit is configured to train the large model based on the emotion evaluation result, the semantic emotion label and the selected training samples to obtain a trained content understanding large model; The content understanding model is used in the method described in claim 1 to obtain sentiment evaluation results and semantic sentiment labels.

19. An electronic device, characterized in that: include: at least one processor; as well as a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method according to any one of claims 1 to 9.

20. A non-transitory computer-readable storage medium storing computer instructions, characterized in that: The computer instructions are used to enable the computer to execute the method according to any one of claims 1 to 9.

21. A computer program product comprising a computer program, wherein when the computer program is executed by a processor, the computer program implements the method according to any one of claims 1 to 9.

Images