Semantic Tag-Based Outbound Voice Quality Inspection Method, System, and Electronic Equipment

Abstract

This invention provides a method, system, and electronic device for outbound voice quality inspection based on semantic tags. The outbound voice quality inspection method includes: in response to receiving an outbound call, acquiring voice data and converting the voice data into dialogue text information; inputting the dialogue text information into a pre-trained semantic tag model to enable the semantic tag model to extract a sequence of semantic tags from the dialogue text information; performing a quality check on the semantic tag sequence based on the quality inspection model to obtain a quality inspection result, wherein the quality inspection model is used to determine whether the semantic tag sequence conforms to the business specifications of outbound call services; and generating a quality inspection report based on the quality inspection result, wherein the quality inspection report includes a quality inspection score and violations. This invention extracts semantic tags from outbound calls through a semantic tag model, achieving accurate semantic parsing of call content, and improves the efficiency and accuracy of outbound voice quality inspection by verifying the semantic tags through a quality inspection model.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a method, system, and electronic device for outbound voice quality inspection based on semantic tags. Background Technology

[0002] In the telecommunications industry, customer service teams need to conduct a large number of outbound marketing calls every day, generating massive amounts of customer dialogue data. The quality of service communication is directly related to customer satisfaction and the sustainable development of the enterprise. Therefore, monitoring the service quality and service process of customer service personnel (i.e., quality inspection) is a key link in enterprise operation and management.

[0003] Currently, mainstream outbound call quality inspection systems fall into two main categories. The first is the traditional manual quality inspection system. This system, based on modules such as recording file management, recording playback control, and manual scoring, supports quality inspectors in developing quality inspection plans, listening to recordings, and scoring online, thus enabling manual quality inspection. The second category is the intelligent quality inspection system, which incorporates AI capabilities, typically including ASR (Automatic Speech Recognition) and intent recognition capabilities. It can perform operations such as transcribing all call recordings, analyzing dialogue text, and automatically scoring, achieving automated quality inspection.

[0004] However, both existing quality inspection systems have significant drawbacks. Traditional manual quality inspection systems require dedicated personnel to listen to recorded calls, identify abnormal emotions, inappropriate language, and sensitive keywords, resulting in high labor costs and significant time consumption. Furthermore, the massive volume of voice recordings makes full quality inspection impossible, requiring only sampling, which leads to the risk of missing many problematic recordings and potentially losing business opportunities. While intelligent quality inspection systems achieve basic intelligent functions, their rules are mostly based on general intents and keywords, failing to provide precise quality evaluation tailored to the specific characteristics of customer service outbound calls in the telecommunications industry. They also suffer from difficulties in accurately recalling calls and system maintenance. Additionally, the rules are demanding to maintain, difficult to learn, lack business orientation, and do not align with the actual quality inspection workflow of customer service outbound calls in the telecommunications industry, thus failing to fully meet the industry's quality inspection needs. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method, system, and electronic device for outbound voice quality inspection based on semantic tags. The aim is to extract semantic tags from the call content through a semantic tag model and use a quality inspection model to perform quality checks on the semantic tags, thereby improving quality inspection efficiency.

[0006] This invention discloses a method for outbound call voice quality inspection based on semantic tags, including: In response to receiving an outbound call, the system acquires the voice data of the outbound call and converts the voice data into dialogue text information, which includes customer service text information and customer text information. The dialogue text information is input into a pre-trained semantic labeling model so that the semantic labeling model can extract a sequence of semantic labels from the dialogue text information. The sequence of semantic labels includes at least one semantic label, and the semantic label is associated with the outbound call business. The quality inspection model is used to perform quality checks on the semantic tag sequence to obtain the quality inspection results. The quality inspection model is used to determine whether the semantic tag sequence conforms to the business specifications of outbound call business. A quality inspection report is generated based on the quality inspection results. The quality inspection report includes the quality inspection score and violations.

[0007] Preferably, converting voice data into dialogue text information includes: Speech recognition technology is used to obtain emotional features and raw text information from speech data; Based on emotion features, the original text information is preprocessed to obtain dialogue text information. The text preprocessing includes semantic enhancement processing and intent correction processing.

[0008] Preferably, semantic tags include general tags, customer interaction tags, and customer service response tags; Customer interaction tags include business inquiry sub-tags, business unsubscription sub-tags, and complaint feedback sub-tags; Customer service response tags include sub-tags for business answers, problem investigation, and complaint handling.

[0009] Preferably, the outbound voice quality inspection method based on semantic tags further includes: Obtain a training sample set, which includes sample dialogue information labeled with semantic tags; The initial semantic labeling model is trained under supervision using a training sample set, so that the initial semantic labeling model learns the mapping relationship between sample dialogue information and semantic labels; The performance of the initial semantic labeling model is evaluated using a pre-built validation set, and the model parameters of the initial semantic labeling model are optimized based on the evaluation results to obtain a pre-trained semantic labeling model.

[0010] Preferably, the semantic tag sequence is subjected to quality inspection based on a quality inspection model to obtain the quality inspection results, including: The semantic label sequence is matched with the preset quality inspection rule base in the quality inspection model. The preset quality inspection rule base contains multiple quality inspection items, each of which includes triggering conditions and scoring logic. If the semantic tag sequence meets the triggering conditions of the quality inspection item, the semantic tag corresponding to the quality inspection item in the semantic tag sequence will be marked as a violation item; The semantic tag sequence is scored according to the scoring logic of the quality inspection items; Based on the violations and quality inspection scores, quality inspection results are generated.

[0011] Preferably, the quality inspection dimensions of the quality inspection model include the standardization of the communication script, service quality, and service attitude; The quality inspection model performs quality checks including single-label quality inspection, multi-label combination quality inspection, and label sequence quality inspection.

[0012] Preferably, after generating a quality inspection report based on the quality inspection results, the method further includes: The quality inspection report and its corresponding dialogue text information are displayed through an interactive interface. In response to user correction actions regarding violations and / or quality inspection scores, generate corrected data; The model parameters of the semantic label model are updated based on the corrected data.

[0013] Preferably, the outbound voice quality inspection method based on semantic tags further includes: Obtain the user's input command to view the violation. The command is used to indicate the location of the text content corresponding to the violation. Query the text content corresponding to the violations in the dialogue text information and highlight the text content.

[0014] This invention discloses a semantic tag-based outbound voice quality inspection system for executing the aforementioned semantic tag-based outbound voice quality inspection method. The outbound voice quality inspection system includes: The data acquisition module is configured to: in response to receiving an outbound call, acquire the voice data of the outbound call and convert the voice data into dialogue text information, including customer service text information and customer text information; The tag generation module is configured to: input dialogue text information into a pre-trained semantic tag model so that the semantic tag model can extract a semantic tag sequence of dialogue text information, the semantic tag sequence including at least one semantic tag, and the semantic tag being associated with outbound call services; The quality inspection module is configured to: perform quality inspection on the semantic tag sequence based on the quality inspection model, and obtain the quality inspection result. The quality inspection model is used to determine whether the semantic tag sequence conforms to the business specifications of outbound call business. The results generation module is configured to generate a quality inspection report based on the quality inspection results. The quality inspection report includes quality inspection scores and violations.

[0015] This invention discloses an electronic 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 semantic tag-based outbound voice quality inspection method.

[0016] Compared with existing technologies, the beneficial effects of this invention are as follows: By introducing a semantic tagging model to accurately extract semantic information from dialogue text, this invention constructs a semantic tagging system tailored to outbound call business scenarios, breaking through the limitations of traditional intelligent quality inspection that relies on general intents and keywords. It can achieve refined quality inspection based on the specific characteristics of outbound call services in the telecommunications industry. The matching mechanism between semantic tag sequences and a preset quality inspection rule base improves the accuracy of quality inspection and reduces the difficulty of system maintenance and the learning curve. Furthermore, through emotion-feedback-assisted text preprocessing, a multi-dimensional quality inspection model, and an iterative mechanism supporting manual correction and feedback, the reliability of the quality inspection results and the adaptability of the system are further enhanced. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the outbound voice quality inspection method based on semantic tags provided by the present invention. Figure 2 This is a schematic diagram of the outbound voice quality inspection system based on semantic tags provided by the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] The present invention will now be described in further detail with reference to the accompanying drawings.

[0020] This invention provides a semantic tag-based outbound voice quality inspection method, applied to a semantic tag-based outbound voice quality inspection system, such as... Figure 1 As shown, the outbound voice quality inspection method based on semantic tags includes the following steps.

[0021] Step S1: In response to receiving an outbound call, obtain the voice data of the outbound call and convert the voice data into dialogue text information.

[0022] In this embodiment of the invention, upon receiving an outbound call, the system captures the voice data during the call in real time, including the voice streams of both the customer service representative and the customer. Subsequently, the voice data is converted into dialogue text information, which includes customer service text information and customer text information.

[0023] In some embodiments, an integrated ASR speech recognition engine transcribes the speech data, converting continuous speech signals into a raw text sequence containing timestamps. Considering potential accent differences, background noise, and speech rate variations in outbound calling scenarios, the ASR engine adaptively optimizes its performance using a professional corpus from the telecommunications industry. For example, it fine-tunes the model for high-frequency business terms such as "data packages" and "broadband renewal" to improve transcription accuracy. After transcription, the system performs role separation on the raw text sequence based on timestamp information, classifying customer service statements as customer service text information and customer statements as customer text information, ultimately forming structured dialogue text information that provides foundational data support for subsequent semantic tag extraction.

[0024] In this embodiment of the invention, speech recognition technology is used to obtain the emotional features and original text information of the speech data; based on the emotional features, the original text information is preprocessed to obtain the dialogue text information. The text preprocessing includes semantic enhancement processing and intent correction processing.

[0025] For example, speech recognition technology, through the synergy of acoustic and language models, first extracts frame-level features from speech data, capturing acoustic features such as fundamental frequency, energy, and Mel-frequency cepstral coefficients in the audio. Then, it combines these with a pre-trained language model to map the acoustic features into the original text information. Simultaneously, it uses an emotion recognition sub-model to analyze emotion-related features such as speech rate, tone, and volume changes, outputting emotion labels for the corresponding sentences, such as "positive," "neutral," "impatient," and "angry." In semantic enhancement processing, the system combines domain dictionaries for outbound calling scenarios (such as "financial products" and "risk assessment" in financial outbound calls, and "logistics timeliness" and "return and exchange policy" in e-commerce outbound calls) to complete ambiguous expressions or colloquial omissions in the original text. For example, it completes the customer service representative's "How do you think about that product?" to "How do you think about the XX model financial product you previously inquired about?" It also standardizes and converts customer dialect words or industry terms, such as converting "This thing is too expensive" to "The price of this product is too high." Intent correction processing calibrates the intent of the original text based on emotional characteristics and contextual information. For example, if a customer expresses "I don't need it right now" with an "impatient" emotional label, the system will consider whether the customer service representative fully explained the product's core benefits in the conversation history to determine if the customer's refusal was due to insufficient information, thus correcting the original intent from "explicit refusal" to "potential hesitation." Similarly, if a customer service representative expresses "Let me check again for you" with a "perfunctory" emotional label, the system will consider whether there is actual inquiry action in subsequent conversations to correct their intent to whether it is a "perfunctory promise." Through this semantic enhancement and intent correction, the system ultimately generates dialogue text information that accurately reflects the content of the conversation and the true state of both parties.

[0026] S2. Input the dialogue text information into the pre-trained semantic labeling model so that the semantic labeling model can extract the semantic label sequence of the dialogue text information.

[0027] In this embodiment of the invention, the semantic tag sequence includes at least one semantic tag, and the semantic tag is associated with the outbound call service. The semantic tags include general tags, customer interaction tags, and customer service response tags. Specifically, the customer interaction tags include service query sub-tags, service unsubscription sub-tags, and complaint feedback sub-tags, and the customer service response tags include service answer sub-tags, problem investigation sub-tags, and complaint handling sub-tags.

[0028] By converting dialogue text into semantic tags, the business logic and interaction details in a call can be captured more accurately, avoiding misjudgments caused by the lack of context in traditional keyword matching. For example, when a customer says, "Why is my data plan from last month being used up so quickly?", the semantic tagging model will first identify the customer's core intent as "business inquiry," and combine it with key information such as "data plan" and "abnormal usage speed" to generate a semantic tag of "customer interaction type - business inquiry sub-tag - data plan usage question." When customer service responds, "Please wait a moment, I will check the specific data consumption details and billing rules within your plan," the model will generate a tag of "customer service response type - problem investigation sub-tag - data consumption details query." These semantic tags are not isolated, but rather form a continuous sequence of semantic tags according to the time sequence of the conversation. They fully record the semantic logic of the entire process from the customer's request to the customer service response and problem handling. For example, the sequence "Customer Interaction - Complaint Feedback Sub-tag - Unresolved Broadband Fault → Customer Service Response - Problem Investigation Sub-tag - Fault Cause Tracing → Customer Service Response - Business Answer Sub-tag - Repair Progress Notification" clearly presents the business flow path in the call.

[0029] During the model training phase, the semantic labeling model constructs a training sample set based on historical call data from the telecommunications industry's outbound calling business. Each sample is labeled with semantically tagged dialogue information. The initial semantic labeling model is then trained under supervision using this training sample set, enabling it to learn the mapping relationship between sample dialogue information and semantic labels. The model employs a BERT pre-trained language model as its foundation. By introducing a dialogue context attention mechanism, the model can simultaneously focus on the correlation between the current statement and the preceding and following dialogue content, improving the coherence and accuracy of label extraction.

[0030] Furthermore, the initial semantic labeling model is evaluated using a pre-built validation set, and its parameters are optimized based on the evaluation results to obtain a pre-trained semantic labeling model. During training, the model is iteratively optimized using the cross-entropy loss function until its label extraction accuracy on the validation set reaches a preset threshold (e.g., above 95%), ensuring that the model can stably output semantic label sequences that fit the business scenario.

[0031] S3. Perform quality checks on the semantic tag sequences based on the quality inspection model to obtain the quality inspection results.

[0032] In this embodiment of the invention, the quality inspection model is used to determine whether the semantic tag sequence conforms to the business specifications of outbound calling services. The quality inspection dimensions of the quality inspection model include the standardization of the call script, service quality, and service attitude; the quality inspection model performs quality checks including single-tag quality inspection, multi-tag combination quality inspection, and tag sequence quality inspection.

[0033] Specifically, single-tag quality inspection independently judges whether a single semantic tag conforms to business specifications. For example, when the semantic tag sequence contains "Customer Service Response - Complaint Handling Sub-tag - Failure to Soothe Customer Emotions," the system will directly trigger the quality inspection rules for the script compliance dimension, determining that the customer service representative did not use soothing language and marking it as a violation. Multi-tag combination quality inspection, on the other hand, combines the relationships between multiple semantic tags for a comprehensive evaluation: if the customer interaction tag is "Complaint Feedback Sub-tag - Broadband Fault Unresolved for One Week," and the customer service response tag is "Business Answer Sub-tag - Only Informed to Wait," the quality inspection model will associate the service quality dimension rules and determine that the customer service representative did not provide a specific solution or repair timeframe, constituting a violation of insufficient service response. Tag sequence quality inspection focuses on whether the temporal logic of semantic tags is reasonable. For example, if the customer interaction tag first appears as "Business Inquiry Sub-tag - Data Package Tariff", the customer service response tag should correspond to "Business Answer Sub-tag - Package Tariff Details". If the answer tag does not appear in the subsequent tag sequence, and instead jumps directly to "Business Promotion Sub-tag - Premium Package Recommendation", the system will determine that the customer service did not follow the "answer first, then recommend" service process, which violates the service quality standards.

[0034] During the quality inspection process, the semantic label sequence is matched against a pre-defined quality inspection rule base in the quality inspection model. This rule base contains multiple quality inspection items, each including triggering conditions and scoring logic. If the semantic label sequence meets the triggering conditions of a quality inspection item, the semantic label corresponding to that item is marked as a violation. The semantic label sequence is then scored according to the scoring logic of the quality inspection item. Based on the violations and the scores, the quality inspection result is generated.

[0035] Specifically, the pre-defined quality inspection rule base sets differentiated triggering conditions and scoring logic for different quality inspection dimensions. If a semantic tag sequence meets the triggering conditions of a quality inspection item, it is marked as a violation, and a quality inspection score is generated.

[0036] For example, the trigger condition for the "timeliness of response" quality inspection item in the service quality dimension is "after the appearance of a customer interaction tag, the corresponding customer service response tag does not appear within 15 seconds". After triggering, points are deducted according to the delay duration: 3 points are deducted for a delay of 15-30 seconds, and 8 points are deducted for a delay of more than 30 seconds.

[0037] For example, the "Problem Resolution Effectiveness" quality inspection item is triggered by a combination of multiple tags. If the customer interaction tag is "Business Request Sub-tag - Abnormal Traffic Deduction", and the "Customer Service Response - Problem Resolution Sub-tag - Explanation of Deduction Reason" or "Customer Service Response - Problem Resolution Sub-tag - Refund Processing Commitment" tag does not appear at the end of the conversation, it is judged as a violation and 15 points will be deducted.

[0038] In some embodiments, when a semantic tag sequence matches the triggering condition of any of the above quality inspection items, the system will automatically mark the corresponding semantic tag as a violation item, and accumulate deductions according to the scoring logic of each item. Finally, the overall quality inspection score (out of 100 points) is calculated by combining the scores of all quality inspection items. At the same time, the system summarizes the type, occurrence time and corresponding dialogue fragments of all violations, and generates a structured quality inspection result containing "violation details", "score details" and "improvement suggestions".

[0039] S4. Generate a quality inspection report based on the quality inspection results.

[0040] In this embodiment of the invention, the quality inspection report includes a quality inspection score and violations.

[0041] For example, the quality inspection report is presented in the form of a structured document, in which the quality inspection score adopts a 100-point scale, directly reflecting the overall compliance level of this outbound call. For example, a score of 92 indicates that the call complies with most business specifications and only has minor wording errors. The violation section is displayed according to the quality inspection dimensions, and each violation includes "violation type", "time stamp of occurrence", "corresponding dialogue segment" and "rule basis".

[0042] In this embodiment of the invention, after generating the quality inspection report, the quality inspection report and its corresponding dialogue text information are displayed on the interactive interface; in response to the user's correction operation on the violation items and / or quality inspection score, correction data is generated; and the model parameters of the semantic tag model are updated based on the correction data.

[0043] Specifically, the system will synchronously display the quality inspection report and corresponding dialogue text information on the interactive interface, allowing users to view the original audio segments and dialogue context corresponding to the violations via a timeline, facilitating quick identification of the problem scenario during manual review. When users disagree with the violations marked by the system, they can submit a manual judgment result through the "Correction" button on the interface.

[0044] The system automatically collects this corrected data and transforms it into training samples with manual annotations. It then updates the parameters of the semantic labeling model through incremental learning: for the corrected semantic labels, the model is retrained to understand the semantics of this type of dialogue scenario; for the adjustment of the scoring logic, the triggering conditions and scoring weights of the corresponding items in the preset quality inspection rule base are optimized simultaneously, so that the model and rule base can continuously adapt to changes in business scenarios and gradually improve the accuracy and intelligence level of quality inspection.

[0045] In this embodiment of the invention, users can locate the text content corresponding to a violation by viewing the violation item. Specifically, the system obtains the user's input instruction to view the violation item, which is used to indicate the location of the text content corresponding to the violation item; it then queries the text information of the dialogue and highlights the text content corresponding to the violation item.

[0046] In this way, users can intuitively find the specific location of the violation in the conversation text, avoiding the tedious process of searching line by line in lengthy text. For example, when a user selects the violation "failure to clearly inform about product risks," the system will immediately locate the sentence segment in the conversation text where the customer service representative did not mention the risk warning, highlight it in red, and at the same time, a floating prompt box will pop up next to the text, displaying the semantic tag type corresponding to the violation and the system's initial judgment basis.

[0047] In some embodiments, if users require further verification, they can click the "Play Corresponding Audio" button associated with the highlighted text to directly listen to the original audio of the dialogue and use voice features such as tone and speed to assist in the judgment. This method of deeply binding violations with text and audio significantly improves the efficiency of manual review and provides a precise reference basis for possible subsequent correction operations.

[0048] like Figure 2 As shown, the present invention also discloses a semantic tag-based outbound voice quality inspection system for executing the aforementioned semantic tag-based outbound voice quality inspection method. The semantic tag-based outbound voice quality inspection system includes: a data acquisition module 201, a tag generation module 202, a quality inspection module 203, and a result generation module 204.

[0049] Specifically, the data acquisition module 201 is configured to: in response to receiving an outbound call, acquire the voice data of the outbound call and convert the voice data into dialogue text information, including customer service text information and customer text information. The tag generation module 202 is configured to: input the dialogue text information into a pre-trained semantic tag model, enabling the semantic tag model to extract a semantic tag sequence from the dialogue text information. The semantic tag sequence includes at least one semantic tag, and the semantic tag is associated with the outbound call service. The quality inspection module 203 is configured to: perform a quality inspection on the semantic tag sequence based on a quality inspection model, obtaining a quality inspection result. The quality inspection model is used to determine whether the semantic tag sequence conforms to the business specifications of the outbound call service. The result generation module 204 is configured to: generate a quality inspection report based on the quality inspection result. The quality inspection report includes a quality inspection score and violations.

[0050] The present invention also discloses an electronic 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 a semantic tag-based outbound voice quality inspection method.

[0051] As can be seen from the above technical solution, this invention provides a method, system, and electronic device for outbound voice quality inspection based on semantic tags. The outbound voice quality inspection method includes: in response to receiving an outbound call, acquiring voice data and converting the voice data into dialogue text information; inputting the dialogue text information into a pre-trained semantic tag model to enable the semantic tag model to extract a semantic tag sequence from the dialogue text information; performing a quality check on the semantic tag sequence based on the quality inspection model to obtain a quality inspection result, wherein the quality inspection model is used to determine whether the semantic tag sequence conforms to the business specifications of outbound call services; and generating a quality inspection report based on the quality inspection result, wherein the quality inspection report includes a quality inspection score and violations. This invention extracts semantic tags from outbound calls through a semantic tag model, achieving accurate semantic parsing of call content, and improves the efficiency and accuracy of outbound voice quality inspection by verifying the semantic tags through a quality inspection model.

[0052] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A semantic tag-based outbound voice quality inspection method, characterized in that, The method comprises the following steps: in response to receiving an outbound call, obtaining voice data of the outbound call, and converting the voice data into conversation text information, the conversation text information comprising customer service text information and customer text information; inputting the conversation text information into a pre-trained semantic label model to enable the semantic label model to extract a semantic label sequence of the conversation text information, the semantic label sequence comprising at least one semantic label, the semantic label being in an associated relationship with an outbound service; performing quality inspection on the semantic label sequence based on a quality inspection model to obtain a quality inspection result, the quality inspection model being used to determine whether the semantic label sequence conforms to a service specification of the outbound service; generating a quality inspection report according to the quality inspection result, the quality inspection report comprising a quality inspection score and a violation item.

2. The outbound voice quality control method of claim 1, wherein, The conversion of the voice data into conversation text information comprises the following steps: obtaining emotion features and original text information of the voice data by using a speech recognition technology; performing text preprocessing on the original text information based on the emotion features to obtain the conversation text information, the text preprocessing comprising semantic enhancement processing and intent correction processing.

3. The outbound voice quality control method of claim 1, wherein, The semantic label comprises a general class label, a customer interaction class label, and a customer service response class label; The customer interaction class label comprises a service query sub-label, a service subscription cancellation sub-label, and a complaint feedback sub-label; The customer service response class label comprises a service answer sub-label, a problem troubleshooting sub-label, and a complaint handling sub-label.

4. The outbound voice quality control method of claim 1, wherein, The method further comprises the following steps: obtaining a training sample set, the training sample set comprising sample conversation information labeled with semantic labels; performing supervised training on an initial semantic label model by using the training sample set to enable the initial semantic label model to learn a mapping relationship between the sample conversation information and the semantic labels; performing performance evaluation on the initial semantic label model by using a pre-constructed verification set, and optimizing model parameters of the initial semantic label model according to an evaluation result to obtain the pre-trained semantic label model.

5. The outbound voice quality control method of claim 1, wherein, The quality inspection on the semantic label sequence based on the quality inspection model to obtain a quality inspection result comprises the following steps: matching the semantic label sequence with a pre-set quality inspection rule library in the quality inspection model, the pre-set quality inspection rule library comprising a plurality of quality inspection items, each of the quality inspection items comprising a trigger condition and a scoring logic; in a case where the semantic label sequence conforms to the trigger condition of the quality inspection item, marking the semantic label corresponding to the quality inspection item in the semantic label sequence as a violation item; performing quality inspection scoring on the semantic label sequence according to the scoring logic of the quality inspection item; generating the quality inspection result based on the violation item and the quality inspection score.

6. The outbound voice quality control method of claim 1, wherein, The quality inspection dimensions of the quality inspection model comprise a dialogue skill specification, a service quality, and a service attitude; The quality inspection model comprises single-label quality inspection, multi-label combination quality inspection, and label sequence quality inspection in the process of performing quality inspection.

7. The outbound voice quality control method of claim 1, wherein, After generating the quality inspection report according to the quality inspection result, the method further comprises the following steps: displaying the quality inspection report and the corresponding conversation text information based on an interaction interface; in response to a user's correction operation on the violation item and / or the quality inspection score, generating correction data; The model parameters of the semantic tag model are updated based on the corrected data. 8.The outbound call voice quality checking method according to claim 7, characterized in that, Also includes: Obtain the user's input instruction to view the violation, wherein the instruction is used to indicate the location of the text content corresponding to the violation; Query the text content corresponding to the violation in the dialogue text information and highlight the text content. 9.A semantic tag based outbound voice quality inspection system, configured to perform the semantic tag based outbound voice quality inspection method according to any one of claims 1-8. include: The data acquisition module is configured to: in response to receiving an outbound call, acquire the voice data of the outbound call, and convert the voice data into dialogue text information, wherein the dialogue text information includes customer service text information and customer text information; The tag generation module is configured to: input the dialogue text information into a pre-trained semantic tag model, so that the semantic tag model extracts a semantic tag sequence of the dialogue text information, wherein the semantic tag sequence includes at least one semantic tag, and the semantic tag is associated with the outbound call service; The quality inspection module is configured to: perform quality inspection on the semantic tag sequence based on the quality inspection model, and obtain the quality inspection result. The quality inspection model is used to determine whether the semantic tag sequence conforms to the business specifications of the outbound call business. The result generation module is configured to generate a quality inspection report based on the quality inspection results, the quality inspection report including quality inspection scores and violations.

10. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the outbound voice quality inspection method based on semantic tags as described in any one of claims 1-8.

Images