Method and device for quality inspection of online customer service speech

Abstract

The present invention discloses a method and device for online customer service speech quality inspection, and relates to the field of computer technology. A specific implementation of the method includes: obtaining online customer service speech quality inspection data, and performing feature extraction to obtain customer sentence vectors and customer service speech vectors; performing customer emotion recognition and intention recognition according to the customer sentence vectors, and performing customer service speech detection according to the customer service speech vectors; performing single sentence verification and dialogue verification according to the emotion recognition results, intention recognition results, and customer service speech detection results, respectively, and performing online customer service speech quality inspection according to the single sentence verification results and the dialogue verification results. This implementation solves the problems of cumbersome manual or rule processing and insufficient role feedback detection, enhances the speech compliance quality inspection capability and accuracy, analyzes the quality inspection speech compliance from multiple angles, effectively saves quality inspection time and improves quality inspection results.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a method and device for quality inspection of online customer service speech. Background Art

[0002] Customer service quality inspection is an important evaluation indicator of customer service performance. The purpose of quality inspection is to improve customer satisfaction, improve customer service, and promote company development and market expansion. Online customer service (different from telephone customer service) plays a vital role in e-commerce and other service fields. Among them, the compliance of speech is the most difficult and most important item in quality inspection. The compliance of speech refers to combining customer characteristics, using positive and active language communication, transforming customer emotions, and creating a good communication atmosphere. However, online customer service quality inspection is difficult because it lacks voice, intonation, tone and other characteristics, and there are few related studies.

[0003] In the process of implementing the present invention, the inventors found that there are at least the following problems in the prior art:

[0004] (1) Online customer service script quality inspection is done manually or by rules, which is cumbersome;

[0005] (2) Insufficient role feedback detection. Checking the compliance of speech is an interactive process, and existing algorithms often separate this interactive process.

[0006] (3) It lacks the ability to analyze from multiple perspectives, and only processes the text in a simple manner without considering the bidirectional characteristics or the contextual relationship of the text. Summary of the invention

[0007] In view of this, an embodiment of the present invention provides a method and device for online customer service speech quality inspection, which can handle various speech compliance quality inspections and solve the tedious problems of manual or rule processing. Among them, the dual-channel aligned input method solves the problem of insufficient role feedback detection and enhances the speech compliance quality inspection capability and accuracy. The device that uses single sentences and dialogues for separate verification can analyze the compliance of quality inspection speech from multiple angles.

[0008] To achieve the above objective, according to one aspect of an embodiment of the present invention, a method for quality inspection of online customer service speech is provided.

[0009] A method for quality inspection of online customer service speech, comprising: obtaining online customer service speech quality inspection data, and performing feature extraction to obtain customer sentence vectors and customer service speech vectors; performing customer emotion recognition and intention recognition based on the customer sentence vectors, and performing customer service speech detection based on the customer service speech vectors; performing single sentence verification and dialogue verification based on emotion recognition results, intention recognition results, and customer service speech detection results, respectively, and performing online customer service speech quality inspection based on the single sentence verification results and the dialogue verification results.

[0010] Optionally, before feature extraction, the method further includes: desensitizing the quality inspection data and replacing key information with characteristic characters; filtering non-text content and robot automatic reply information; and dividing conversations according to the customer's consultation content.

[0011] Optionally, before performing feature extraction, it also includes: dividing the online customer service speech quality inspection data into customer statements and customer service speeches according to the conversation subject; performing feature extraction to obtain customer statement vectors and customer service speech vectors includes: using a pre-trained model to perform feature extraction on the customer statements and the customer service speech respectively to obtain customer statement vectors and customer service speech vectors.

[0012] Optionally, after using the pre-trained model to extract features of the customer service speech to obtain a customer service speech vector, the method further includes: evaluating the pre-training effect of the customer service speech vector using a cosine similarity method between vectors, and adjusting parameters of the pre-trained model.

[0013] Optionally, before performing customer emotion recognition and intention recognition based on the customer sentence vector and detecting customer service speech based on the customer service speech vector, the method further includes: annotating the quality inspection data for offensive language; and marking the customer sentence vector and the customer service speech vector for roles and positions.

[0014] Optionally, identifying customer emotions based on the customer sentence vector includes: bidirectionally inputting the customer sentence vector into an emotion recognition model to identify customer emotions, wherein the emotion recognition model uses a gated recurrent unit to filter out emotion-irrelevant vectors, and then accesses a normalization layer to output customer emotions.

[0015] Optionally, identifying customer intention based on the customer sentence vector includes: obtaining discrete features for representing customer intention based on order information; inputting the discrete features and the customer sentence vector into an intention recognition model to identify customer intention, wherein the intention recognition model identifies customer intention in the following manner: for the discrete features, high-dimensional mapping is used followed by vector dot multiplication for feature crossover to obtain a first input; for the customer speech vector, a convolutional neural network is used for encoding and connected to a fully connected layer, and the fully connected layer is crossovered with the features to obtain a second input; the first input and the second input are concatenated and connected to a classification layer to obtain an intention recognition result.

[0016] Optionally, performing customer service speech detection based on the customer service speech vector includes: inputting the customer service speech vector into a fine-tuning layer to fine-tune the customer service speech vector; connecting the fine-tuning result to a logarithmic loss function for binary classification to obtain a customer service speech detection result.

[0017] Optionally, performing single sentence verification based on emotion recognition results, intention recognition results and customer service speech detection results includes: inputting the emotion recognition results, the intention recognition results, the customer service speech detection results and the customer service speech vector corresponding to the single sentence customer service speech, and the customer sentence vector corresponding to the context customer speech of the single sentence customer service speech into a single sentence verification model for single sentence verification, wherein the single sentence verification model performs one-hot coding mapping on the emotion recognition results, the intention recognition results and the customer service speech detection results to obtain a one-hot code vector, establishes a gating unit with the one-hot code vector, and sets trainable parameters to establish an association, filters the input content, and uses a multi-head attention mechanism to perform multi-angle attention cross-attention on the filtered content to perform single sentence verification.

[0018] Optionally, performing conversation verification based on emotion recognition results, intention recognition results and customer service speech detection results includes: inputting the customer sentence vector and customer service speech vector, emotion recognition results, intention recognition results, and customer service speech detection results corresponding to the complete conversation into a conversation verification model for conversation verification, wherein the conversation verification model performs one-hot coding mapping on the emotion recognition results, the intention recognition results, and the customer service speech detection results to obtain a one-hot code vector, establishes a gating unit with the one-hot code vector, and sets trainable parameters to establish an association, filters the input content, and uses a multi-head attention mechanism to perform multi-angle attention cross-attention on the filtered content for conversation verification.

[0019] Optionally, performing online customer service speech quality inspection based on the single sentence verification result and the dialogue verification result includes: performing weighted averaging of the single sentence verification result and the dialogue verification result to obtain a quality inspection result for performing online customer service speech quality inspection.

[0020] Optionally, the method also includes: identifying positive rules and negative rules on the customer service speech quality inspection data; obtaining customer service operation records and performing one-hot code mapping on the operation records to obtain a one-hot code vector, and concatenating the one-hot code vector with the rule recognition result to obtain a first quality inspection result; merging the first quality inspection result with the quality inspection result, and matching different weights to obtain a final quality inspection result.

[0021] According to another aspect of an embodiment of the present invention, a device for quality inspection of online customer service speech is provided.

[0022] A device for online customer service speech quality inspection, comprising: a data acquisition module, used to acquire online customer service speech quality inspection data, and perform feature extraction to obtain customer sentence vectors and customer service speech vectors; an intelligent recognition module, used to perform customer emotion recognition and intention recognition based on the customer sentence vectors, and perform customer service speech detection based on the customer service speech vectors; a speech quality inspection module, used to perform single sentence verification and dialogue verification based on emotion recognition results, intention recognition results and customer service speech detection results, respectively, and perform online customer service speech quality inspection based on the single sentence verification results and the dialogue verification results.

[0023] Optionally, the device also includes a data processing module, which is used to: desensitize the quality inspection data and replace key information with characteristic characters; filter non-text content and robot automatic reply information; and divide conversations according to the customer's consultation content.

[0024] Optionally, the device also includes a vector acquisition module, which is used to: divide the online customer service speech quality inspection data into customer statements and customer service speech according to the conversation subject; perform feature extraction to obtain customer statement vectors and customer service speech vectors, including: use a pre-trained model to perform feature extraction on the customer statements and the customer service speech respectively to obtain customer statement vectors and customer service speech vectors.

[0025] Optionally, the device further includes a parameter adjustment module, which is used to: evaluate the pre-training effect of the customer service speech vector using a cosine similarity method between vectors, and adjust the parameters of the pre-training model.

[0026] Optionally, the device further comprises a data marking module, which is used to: mark the quality inspection data for offensive language; and mark the customer sentence vector and the customer service speech vector for roles and positions.

[0027] Optionally, the intelligent recognition module is also used to: bidirectionally input the customer sentence vector into an emotion recognition model to perform customer emotion recognition, wherein the emotion recognition model uses a gated recurrent unit to filter out emotion-irrelevant vectors, and then accesses a normalization layer to output customer emotions.

[0028] Optionally, the intelligent recognition module is also used to: obtain discrete features for representing customer intentions based on customer statements; input the discrete features and the customer statement vector into an intention recognition model to recognize customer intentions, wherein the intention recognition model recognizes customer intentions in the following manner: for the discrete features, high-dimensional mapping is used followed by vector dot multiplication for feature crossover to obtain a first input; for the customer speech vector, a convolutional neural network is used for encoding and connected to a fully connected layer, and the fully connected layer is crossovered with the features to obtain a second input; the first input and the second input are concatenated and connected to a classification layer to obtain an intention recognition result.

[0029] Optionally, the intelligent recognition module is also used to: input the customer service speech vector into the fine-tuning layer to fine-tune the customer service speech vector; connect the fine-tuning result to the logarithmic loss function for binary classification to obtain the customer service speech detection result.

[0030] Optionally, the speech quality inspection module is also used to: input the emotion recognition result, the intention recognition result, the customer service speech detection result, the customer service speech vector corresponding to the single-sentence customer service speech, and the customer sentence vector corresponding to the context customer speech of the single-sentence customer service speech into a single sentence verification model for single sentence verification, wherein the single sentence verification model performs one-hot coding mapping on the emotion recognition result, the intention recognition result, and the customer service speech detection result to obtain a one-hot code vector, establishes a gating unit with the one-hot code vector, and sets trainable parameters to establish an association, filters the input content, and uses a multi-head attention mechanism to perform multi-angle attention cross-attention on the filtered content for single sentence verification.

[0031] Optionally, the speech quality inspection module is also used to: input the customer sentence vector and customer service speech vector, emotion recognition results, intent recognition results, and customer service speech detection results corresponding to the complete conversation into a dialogue verification model for dialogue verification, wherein the dialogue verification model performs one-hot coding mapping on the emotion recognition results, the intent recognition results, and the customer service speech detection results to obtain a one-hot code vector, establishes a gating unit with the one-hot code vector, and sets trainable parameters to establish an association, filters the input content, and uses a multi-head attention mechanism to perform multi-angle attention cross-attention on the filtered content for dialogue verification.

[0032] Optionally, the speech quality inspection module is also used to: perform weighted averaging of the single sentence verification result and the conversation verification result to obtain a quality inspection result for performing online customer service speech quality inspection.

[0033] Optionally, the device also includes a final result module, which is used to: identify positive rules and negative rules on the customer service speech quality inspection data; obtain customer service operation records and perform one-hot code mapping on the operation records to obtain a one-hot code vector, and concatenate the one-hot code vector with the rule recognition result to obtain a first quality inspection result; merge the first quality inspection result with the quality inspection result, and match different weights to obtain a final quality inspection result.

[0034] According to yet another aspect of an embodiment of the present invention, an electronic device for quality inspection of online customer service speech is provided.

[0035] An electronic device for online customer service speech quality inspection, comprising: one or more processors; a storage device for storing one or more programs, when the one or more programs are executed by the one or more processors, the one or more processors implement the method for online customer service speech quality inspection provided by an embodiment of the present invention.

[0036] According to still another aspect of the embodiments of the present invention, a computer readable medium is provided.

[0037] A computer-readable medium stores a computer program, which, when executed by a processor, implements the method for quality inspection of online customer service speech provided by an embodiment of the present invention.

[0038] An embodiment of the above invention has the following advantages or beneficial effects: obtaining online customer service speech quality inspection data and performing feature extraction to obtain customer sentence vectors and customer service speech vectors; performing customer emotion recognition and intention recognition based on the customer sentence vectors, and performing customer service speech detection based on the customer service speech vectors; performing single sentence verification and dialogue verification based on the emotion recognition results, intention recognition results and customer service speech detection results respectively, and performing technical means for online customer service speech quality inspection based on the single sentence verification results and the dialogue verification results, which can handle a variety of speech compliance quality inspections and solve the tedious problems of manual or rule processing. The dual-channel aligned input method solves the problem of insufficient role feedback detection and enhances the speech compliance quality inspection capability and accuracy. The device for separate verification of single sentences and dialogues can analyze the quality inspection speech compliance from multiple angles, and effectively saves quality inspection time and improves quality inspection results through diversified customer service quality inspection models divided by roles and types.

[0039] The further effects of the above-mentioned non-conventional optional manner will be described below in conjunction with the specific implementation manner. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The accompanying drawings are used to better understand the present invention and do not constitute an improper limitation of the present invention.

[0041] Figure 1This is a schematic diagram of the main steps of a method for quality inspection of online customer service speech according to an embodiment of the present invention;

[0042] Figure 2 This is a schematic diagram of the main process of the method for quality inspection of online customer service speech according to an embodiment of the present invention;

[0043] Figure 3 is a schematic diagram of the main process of intent recognition according to an embodiment of the present invention;

[0044] Figure 4 is a schematic diagram of main modules of an apparatus for online customer service speech quality inspection according to an embodiment of the present invention;

[0045] Figure 5 is an exemplary system architecture diagram to which embodiments of the present invention may be applied;

[0046] Figure 6 It is a schematic diagram of the structure of a computer system of a terminal device or a server suitable for implementing an embodiment of the present invention. DETAILED DESCRIPTION

[0047] The following is a description of exemplary embodiments of the present invention in conjunction with the accompanying drawings, including various details of the embodiments of the present invention to facilitate understanding, which should be considered as merely exemplary. Therefore, it should be recognized by those of ordinary skill in the art that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present invention. Similarly, for clarity and conciseness, the description of well-known functions and structures is omitted in the following description.

[0048] At present, the main online customer service script compliance-related quality inspection methods are as follows:

[0049] (1) The manual quality inspection method mainly relies on manual review. The inspection items of customer service quality inspection generally include opening remarks, closing remarks, service specifications, service soft skills, etc. Customer service quality inspection usually randomly selects 5% from the chat conversations for quality inspection by professional quality inspectors, and then re-selects 5% from the first sampling results for re-inspection as the final result. If an appeal is filed, additional quality inspection is required;

[0050] (2) Rule-based quality inspection method mainly defines key words and then detects whether key words or sentence structures are used in online chat records. For example, standard words are sent to politely greet consulting customers, such as: Hello, customer service ### employee number is happy to serve you, or certain prohibited words are defined, such as avoiding the use of weak words in communication with customers, such as "I can't...", "We don't...", "Our policy states that...", "I don't know...", etc., avoiding colloquialisms, daily life, and catchphrases, etc., and obtaining compliance results through matching of forward rules and reverse rules;

[0051] (3) Intelligent quality inspection methods: one is to use traditional machine learning algorithms, such as support vector machine (SVM, a classification technology) to classify the converted vectors of text input; the other is to use the pre-research model in deep learning. The main process includes: data labeling, data preprocessing, feature encoding (including one hot (word2vec (related model used to generate word vectors))) + position encoding, n-gram (an algorithm based on statistical language model)), classification model selection (recurrent neural network, convolutional neural network) and other steps.

[0052] The inventors analyzed the existing technical solutions and found that the above-mentioned customer service speech compliance-related quality inspection methods have the following shortcomings:

[0053] (1) Manual or rule-based processing is cumbersome

[0054] Manual processing or rule-based quality inspection is too cumbersome. Manual quality inspection requires at least three rounds of inspection, and each round of inspection requires listening to the complete phone recording and making a judgment. This is because customer service quality inspection is often directly related to customer service performance and must be taken seriously. Although the rule-based method can assist in the quality inspection of the compliance of the speech to a certain extent, the process of setting rules is a relatively cumbersome and complicated task and the applicable scenarios are extremely limited, so it can only provide a certain degree of assistance;

[0055] (2) Insufficient role feedback detection

[0056] Compliance checking of speech is often an interactive process, such as the dialogue shown in Table 1 below:

[0057] Table 1

[0058]

[0059] In terms of whether the customer service staff calms the customer's emotions, we need to pay attention not only to the customer service staff's behavior, but also to whether the customer's emotions change under the guidance of the customer service staff. In the example in the table above, the customer expressed his dissatisfaction and anger with the handling result. The customer service staff calmed the customer's emotions by promptly guiding the customer's emotions and proposing acceptable solutions. In the next paragraph, the customer's emotions obviously improved. Therefore, this kind of compliance check of speech is not only from the perspective of customer service, but also a process of interactive feedback between different roles. The existing algorithm method often separates this feedback process, or only inputs the customer service corpus or directly inputs it without distinction, which leads to extremely poor model detection effect and cannot be truly applied in the field of quality inspection;

[0060] (3) Lack of multi-angle analysis capabilities

[0061] The so-called multi-angle analysis capability refers to analyzing the compliance of quality inspection words from the perspective of dialogue and single sentence. A typical example is shown in Table 2 below:

[0062] Table 2

[0063]

[0064] In the example in the table above, if we look at it from a single sentence perspective, it is very easy to interpret "I'm very sorry, the leader cannot connect, it's very inconsistent" as shirking responsibility. However, from a conversational perspective, we can see from the following sentence supplement "But I've applied to the leader for you, and I'll call you later to give you feedback. Is that okay?" that the customer service staff is actively calming the customer's emotions and meeting their requirements, which has improved the customer's emotions. Therefore, it is also necessary to analyze the conversational perspective in context. We cannot separate the quality inspection content, but the existing intelligent algorithms only process the text simply, without considering the bidirectional characteristics, let alone the text contextual relationship. The model's ability to distinguish between single sentences and conversational perspectives is necessary.

[0065] The technical solution of the present invention provides a multi-channel intelligent algorithm solution for the problem of cumbersome manual or rule processing. On the one hand, strong rules are used to automatically identify directly identifiable problems. On the other hand, intelligent analysis of text content and application of deep learning algorithms can achieve the purpose of automatic review of compliance quality inspection of various speech techniques. In extreme cases, quality inspectors only need to conduct a round of 5% random inspection and re-inspection. In view of the problem of insufficient role feedback detection, the present invention provides a dual-channel aligned input method. The emotion recognition module in the method can effectively detect the emotional changes of customers and thus assist in proving the compliance of customer service speech techniques. At the same time, the method also judges whether the customer service effectively identifies the customer problem and provides a solution based on the degree of matching between the customer service and the customer text questions and answers. This two-way verification feedback method enhances the speech compliance quality inspection capability and accuracy. In view of the lack of multi-angle analysis capability, the present invention uses a device for separate verification of single sentences and dialogues. The single sentence verification device mainly checks whether the customer service's single sentence has obvious non-compliant speech communication and ignores the detected customer intention. The dialogue verification device mainly checks whether the customer service's continuous dialogue follows the principle of good communication and reasonably guides and establishes solutions for customer requirements.

[0066] Figure 1 Schematic diagram of the main steps of the method for quality inspection of online customer service speech according to an embodiment of the present invention. Figure 1 As shown, the method for online customer service speech quality inspection in an embodiment of the present invention mainly includes the following steps S101 to S103.

[0067] Step S101, obtain online customer service speech quality inspection data, and perform feature extraction to obtain customer sentence vectors and customer service speech vectors.

[0068] In an embodiment of the present invention, before feature extraction, it also includes: desensitizing the quality inspection data, replacing key information with characteristic characters; filtering non-text content and robot automatic reply information; dividing the conversation according to the customer's consultation content. In the present invention, the compliance check of online customers' speech in the e-commerce field is taken as an example. The quality inspection data does not have features such as timbre and intonation, but the accuracy of the text is guaranteed, and there is no problem of homophone disambiguation in voice transcription. The quality inspection data needs to be desensitized. For the e-commerce field, key information includes order number, customer PIN, waybill number, and mobile phone number, which need to be replaced with characteristic characters. After that, non-text content needs to be filtered out, such as picture links and some garbled English letters, etc. Regular expressions can be used for filtering, and the processed data is passed to the dialogue processing layer. In the dialogue processing layer, dialogue verification is mainly performed. This is because the model The model needs to accurately identify customer service and customer role information in order to improve the accuracy of the algorithm. First, the robot's automatic reply information is filtered. When the robot is activated, the reply information can be accurately labeled and can be filtered directly. Secondly, the complete conversation is checked and divided. This is because the customer and customer service may consult the same event again after a period of time during the communication process. The main identification method is to match it through the consultation item selected by the customer. At the same time, in each consultation, the previous item of the last consultation is found for context matching. If the similarity reaches the threshold, it is regarded as the same event consultation. Since the customer and customer service roles are online consultations, 100% accuracy can also be guaranteed.

[0069] In an embodiment of the present invention, before feature extraction, it also includes: dividing the online customer service speech quality inspection data into customer sentences and customer service speeches according to the dialogue subject; feature extraction to obtain customer sentence vectors and customer service speech vectors includes: using pre-trained models to extract features from customer sentences and customer service speeches respectively to obtain customer sentence vectors and customer service speech vectors. In the present invention, the pre-trained model layer is used to achieve the purpose of obtaining customer service and customer vectors. The main function of this layer is to convert the dialogue text between customer service and customers into a vector matrix as a subsequent model input. In this layer, the ALBERT model is used to pre-train the data. The ALBERT model is lightweight and easy to deploy and has stable online effects. At the same time, the Transformer+Attention model has good applicability to contextual information and short dialogues. The customer and customer service roles are pre-trained to obtain customer sentence vectors and customer service speech vectors.

[0070] According to one embodiment of the present invention, after using a pre-trained model to extract features of the customer service speech to obtain a customer service speech vector, it also includes: for the customer service speech vector, using the cosine similarity method between vectors to evaluate the pre-training effect, adjusting the parameters of the pre-training model, the customer service language has good uniformity, for the customer service language vector, using the cosine similarity method between vectors to evaluate the pre-training effect, and adjusting the parameters of the pre-training model.

[0071] According to another embodiment of the present invention, before performing customer emotion recognition and intent recognition based on the customer sentence vector and performing customer service speech detection based on the customer service speech vector, it also includes: marking offensive language on the quality inspection data; marking the roles and positions of the customer sentence vector and the customer service speech vector. Customer service and customer language are sometimes aggressive. To address this problem, you can use manual methods to strongly mark offensive language, such as abusive language, that is, strong dissatisfaction. When processing the customer service speech vector and the customer sentence vector, you will additionally add the role and position mark encoding to the pre-training data, and then pass the customer vector matrix to the emotion recognition layer and the intent recognition layer, and pass the customer service vector matrix to the speech detection layer for the next steps.

[0072] Step S102, performing customer emotion recognition and intention recognition based on the customer sentence vector, and performing customer service speech detection based on the customer service speech vector.

[0073] In an embodiment of the present invention, the emotion recognition of a customer based on a customer sentence vector includes: bidirectionally inputting the customer sentence vector into an emotion recognition model to perform emotion recognition on the customer, wherein the emotion recognition model uses a gated recurrent unit to filter out emotion-irrelevant vectors, and then accesses a normalization layer to output the customer's emotion. The emotion recognition of the customer is achieved through the emotion recognition layer. Actual data research shows that when a customer expresses emotions, a complete sentence represents an emotion, and it is usually a short text. Therefore, only coarse-grained, i.e., whole-sentence sentiment analysis is required, and there is no need to extract evaluation words. The method used here is to bidirectionally input the pre-training results of the text into the model, use a gated recurrent unit in the model to filter out emotion-irrelevant vectors, and finally access a normalization layer to output the final emotion judgment.

[0074] According to another embodiment of the present invention, identifying customer intentions based on customer sentence vectors includes: obtaining discrete features for representing customer intentions based on customer sentences; inputting the discrete features and the customer sentence vectors into an intention recognition model to identify customer intentions, wherein the intention recognition model identifies customer intentions in the following manner: for discrete features, high-dimensional mapping is used followed by vector dot multiplication for feature crossover to obtain a first input; for customer speech vectors, a convolutional neural network is used for encoding and connected to a fully connected layer, and the fully connected layer is crossovered with the features to obtain a second input; the first input and the second input are concatenated and connected to a classification layer to obtain an intention recognition result.

[0075] Figure 3 FIG. 1 is a schematic diagram of the main process of intent recognition according to an embodiment of the present invention. Figure 3 As shown in the figure, customer intent recognition is achieved through the intent recognition layer. The intent recognition layer refers to summarizing a specific demand from the customer's consultation questions. For example, "Hello, the mobile phone I bought yesterday cannot be charged. How can a new mobile phone have such a problem? Please return it to me as soon as possible." The intent recognition model can identify the intention of "return". The intent recognition layer needs to introduce other additional supplementary features. Customer consultation questions are usually not simply judged by the customer's own words. Other dimensional features of the customer can be matched according to the customer's consultation order, such as return and exchange orders, dispute orders, or direct order information. This feature can be directly matched and searched from the options when the customer consults the question, and these features are regarded as discrete features and input into the intent recognition model together with the dialogue. For discrete features, high-dimensional mapping is used, and vector dot multiplication is used for feature crossover to automatically mine additional features. The text is encoded using a convolutional neural network and connected to the fully connected layer, and the fully connected layer is spliced ​​with the feature crossover, and finally connected to the classification layer to obtain the intent recognition conclusion.

[0076] In an embodiment of the present invention, customer service speech detection based on the customer service speech vector includes: inputting the customer service speech vector into the fine-tuning layer to fine-tune the customer service speech vector; connecting the fine-tuning result to the logarithmic loss function for binary classification to obtain the customer service speech detection result. The customer service speech detection is implemented through the speech detection layer, and the language model is used to synchronously check the standardization of the communication speech, which can be regarded as a binary classification problem. The pre-trained vector is then connected to the fine-tuning layer. The fine-tuning layer can perform some high-dimensional model dimension analysis on the input vector to obtain better results during classification, and then the pre-training result is fine-tuned and connected to the log loss function for binary classification.

[0077] Step S103, performing single sentence verification and dialogue verification respectively according to the emotion recognition results, intention recognition results and customer service speech detection results, and performing online customer service speech quality inspection according to the single sentence verification results and the dialogue verification results.

[0078] In an embodiment of the present invention, single sentence verification is performed based on emotion recognition results, intention recognition results and customer service speech detection results, including: inputting the emotion recognition results, intention recognition results, customer service speech detection results and the customer service speech vector corresponding to the single sentence customer service speech, and the customer sentence vector corresponding to the context customer speech of the single sentence customer service speech into a single sentence verification model for single sentence verification, wherein the single sentence verification model performs one-hot coding mapping on the emotion recognition results, intention recognition results and customer service speech detection results to obtain a one-hot code vector, establishes a gating unit with the one-hot code vector, and sets trainable parameters to establish association, filters the input content, and uses a multi-head attention mechanism to perform multi-angle attention cross-attention on the filtered content to perform single sentence verification. Single sentence verification is achieved through the single sentence verification layer. The single sentence verification layer first starts from the single sentence of the customer service. The single sentence verification layer is a cross layer. The input is the customer emotion, customer intention and customer service speech detection results, the customer service speech vector corresponding to the single sentence customer service speech, and the customer sentence vector corresponding to the context customer speech of the single sentence customer service speech. The above recognition content is one-hot mapped to obtain a one-hot vector. At the same time, the context customer vector and recognition content including the customer service reply content are input. This is to ensure that the model can calculate the complete processing process of the customer service response to the customer intention or emotion. By establishing a gating unit with the one-hot vector, the customer sentence vector and the customer service speech vector and setting trainable parameters to establish an association, the input content is filtered, and then the filtered content is cross-attended, and a multi-head attention mechanism is used to cross-attack at multiple angles. This method can effectively test the impact of customer service speech on the changes in customer emotions before and after, and the attention cross-attack can effectively test the degree of customer service's processing of customer intentions.

[0079] In an embodiment of the present invention, performing dialogue verification based on emotion recognition results, intention recognition results, and customer service speech detection results includes: inputting the customer sentence vector and customer service speech vector, emotion recognition results, intention recognition results, and customer service speech detection results corresponding to the complete conversation into a dialogue verification model for dialogue verification, wherein the dialogue verification model performs one-hot code mapping on the emotion recognition results, intention recognition results, and customer service speech detection results to obtain a one-hot code vector, establishes a gating unit with the one-hot code vector and the customer sentence vector and the customer service speech vector and sets trainable parameters to establish an association, filters the input content, and uses a multi-head attention mechanism to perform attention cross-over from multiple angles on the filtered content for dialogue verification. The dialogue verification is implemented through a dialogue verification layer, which starts from the complete dialogue of the customer service. The device setting is similar to a single-sentence device, and all customer service speech vectors and customer sentence vectors are input into the dialogue verification model according to role interaction, and the content is cross-overed using multi-head attention. This method can avoid the one-sided quality inspection results caused by single-sentence judgment.

[0080] In an embodiment of the present invention, performing online customer service speech quality inspection based on single sentence verification results and dialogue verification results includes: performing weighted average of single sentence verification results and dialogue verification results to obtain quality inspection results for online customer service speech quality inspection. Single sentence verification and dialogue verification results are output through a multi-mode output layer, and the multi-mode output layer performs weighted average of single sentence verification devices and dialogue verification devices to output a final conclusion, while also retaining the results of the sentences, providing a fine-grained speech compliance quality inspection effect.

[0081] In an embodiment of the present invention, the method also includes: identifying positive rules and negative rules for customer service speech quality inspection data; obtaining customer service operation records and mapping the operation records to obtain a one-hot code vector, splicing the one-hot code vector with the rule recognition result to obtain a first quality inspection result; merging the first quality inspection result with the quality inspection result, matching different weights to obtain a final quality inspection result. In the specific application process, rule identification is implemented through a strong rule matching layer. In this layer, we still retain some strong speech rules for speech compliance. This rule is 100% accurate and is divided into positive rules and negative rules. Positive rules use honorifics such as "you", "please", and "thank you". Negative rules use colloquial words or weak words such as "bye", "hmm?", and "ah?". This rule has a small recognition range, but its accuracy is guaranteed. It is also an essential auxiliary detection method in quality inspection, and it complements the deep learning model. At the same time, in order to detect whether the customer service accurately responds to the customer's intention, it is necessary to check the customer service's operation records. This is mainly analyzed from the customer service click records in the customer service system. For example, if the customer service promises to apply for coupon compensation, it is determined whether the customer service has performed the coupon compensation operation. The operation record is one-hot mapped, and then spliced ​​with the result of strong rule matching, and input into the joint verification layer together with the result of the multi-mode output layer. For the model, whether the result of intent recognition is related to the vector mapping of the customer service operation record, the model can learn automatically. The joint verification layer combines the multi-mode output result with the splicing result of the strong rule matching layer. The multi-mode output result merges the result of single sentence verification and the result of dialogue verification with the splicing result of the strong rule. Different weights are assigned to the two according to the strength of the rule setting, but the rule weight is generally set higher overall. For data that the strong rule has no discrimination ability, since the multi-mode output layer is a log function, it can display the compliance probability value, and the content that the model considers to have weak discrimination can be output to facilitate the final re-inspection by the quality inspector.

[0082] Figure 2 FIG. 1 is a schematic diagram of the main flow of a method for quality inspection of online customer service speech according to an embodiment of the present invention. Figure 2 As shown in the figure, the main process of the online customer service speech quality inspection method includes:

[0083] (1) The quality inspection data layer obtains the quality inspection data, desensitizes the quality inspection data, filters non-text content, and transmits the processed quality inspection data to the dialogue processing layer;

[0084] (2) The dialogue processing layer is mainly responsible for dialogue verification, first filtering the robot's automatic reply information, and secondly performing complete conversation verification;

[0085] (3) Strong rule matching layer: input the data from the dialogue processing layer into the strong rule matching layer to identify positive and negative rules, obtain customer service operation records, perform one-hot mapping on the customer service operation records, and match the mapping results with the rule identification results;

[0086] (4) Model pre-training layer: The model pre-training layer converts the conversation texts of customer service and customers into vectors as subsequent inputs. Since customer service language has good uniformity, the cosine similarity method between vectors is used to evaluate the pre-training effect and adjust the pre-training parameters. The language of customer service and customers is sometimes aggressive. We use manual methods to strongly label aggressive language, such as abusive language, which means strong dissatisfaction.

[0087] (5) Training data preparation layer: The customer sentence vector and customer service speech vector converted from the conversation text between the customer and customer service roles are respectively passed to the training data preparation layer. This layer is responsible for adding the role and position tag encoding to the vector, and passing the customer sentence vector to the emotion recognition layer and the intention recognition layer, and passing the customer service speech vector to the speech detection layer;

[0088] (6) Emotion recognition layer: The customer sentence vector is bidirectionally input into the emotion recognition layer model. A gated recurrent unit is used in the model to filter out emotion-irrelevant vectors. Finally, the normalization layer is connected to output the final emotion discrimination.

[0089] (7) Intent recognition layer: The customer sentence vector is input into the intent recognition layer to summarize a specific demand from the customer's consultation question and identify the customer's intent;

[0090] (8) Speech detection layer: The customer service speech vector output by the model pre-training layer is input into the speech detection layer, which is mainly used to check the standardization of communication speech;

[0091] (9) Single sentence verification layer and dialogue verification layer. The single sentence verification layer is a cross layer. It starts from a single sentence of the customer service and inputs customer emotions, customer intentions, speech detection results, customer service speech vectors, and corresponding context customer vectors. The dialogue verification layer starts from the complete dialogue of the customer service. The dialogue verification layer is similar to the single sentence verification layer. All customer sentence vectors and customer service speech vectors are input into the dialogue verification layer according to role interactions. This effectively tests the impact of customer service speech on customer emotions from multiple angles.

[0092] (10) The multi-modal output layer performs a weighted average of the results of the single sentence verification layer and the dialogue verification layer to output the final conclusion. It also retains the results of the sentence divisions and provides fine-grained compliance inspection results for the speech.

[0093] (11) Joint verification layer: The main function of this layer is to combine the multi-mode output results with the splicing results of the strong rule matching layer, and to assign different weights according to the strength of the rule setting to obtain the final quality inspection results.

[0094] The compliance of customer service speech is the most critical item in customer service quality inspection, and it is a huge challenge for online customer service. The method proposed in the present invention, which cleverly couples deep learning, rule matching and manual re-inspection, can improve the efficiency of detection and ensure the correctness of the detection results. The dual-channel aligned input method can distinguish between customer and customer service roles for quality inspection, which includes multiple shared layers such as intention recognition, emotion recognition, and speech detection, and can effectively detect the effect of customer service speech based on context information. The method of separately verifying single sentences and dialogues can effectively apply customer changes to quality inspection by understanding different context information. Through the filtering device of text intersection and gating unit, the speech score is analyzed from the perspective of single sentence and dialogue respectively and the probability is output. The diversified customer service quality inspection model based on role and type can effectively save quality inspection time and improve the effect. In the online quality inspection in the e-commerce field, after adjusting the parameters, the overall accuracy can reach 95.78% without changing the distribution of quality inspection samples. The time for a round of quality inspection can be completely omitted, and the re-inspection time can be saved by more than 50%.

[0095] Figure 4 is a schematic diagram of the main modules of the device for online customer service speech quality inspection according to an embodiment of the present invention, such as Figure 4 As shown, the device 400 for online customer service speech quality inspection according to an embodiment of the present invention mainly includes a data acquisition module 402 , an intelligent recognition module 403 , and a speech quality inspection module 403 .

[0096] The data acquisition module 401 is used to acquire online customer service speech quality inspection data and perform feature extraction to obtain customer sentence vectors and customer service speech vectors;

[0097] Intelligent recognition module 402, used to perform customer emotion recognition and intention recognition based on customer sentence vectors, and to perform customer service speech detection based on customer service speech vectors;

[0098] The speech quality inspection module 403 is used to perform single sentence verification and dialogue verification according to the emotion recognition results, intention recognition results and customer service speech detection results, and perform online customer service speech quality inspection according to the single sentence verification results and the dialogue verification results.

[0099] In an embodiment of the present invention, the device 400 also includes a data processing module (not shown in the figure), which is used to: desensitize the quality inspection data and replace key information with characteristic characters; filter non-text content and robot automatic reply information; and divide the conversation according to the customer's consultation content.

[0100] In an embodiment of the present invention, the device 400 also includes a vector acquisition module (not shown in the figure), which is used to: divide the online customer service speech quality inspection data into customer sentences and customer service speeches according to the dialogue subject; perform feature extraction to obtain customer sentence vectors and customer service speech vectors, including: use pre-trained models to perform feature extraction on customer sentences and customer service speeches respectively to obtain customer sentence vectors and customer service speech vectors.

[0101] In an embodiment of the present invention, the device 400 also includes a parameter adjustment module (not shown in the figure), which is used to: evaluate the pre-training effect of the customer service speech vector using the cosine similarity method between vectors, and adjust the parameters of the pre-training model.

[0102] In an embodiment of the present invention, the device 400 further includes a data labeling module (not shown in the figure) for: marking offensive language on quality inspection data; and labeling roles and positions on customer sentence vectors and customer service speech vectors.

[0103] In an embodiment of the present invention, the intelligent recognition module 402 is also used to: bidirectionally input the customer sentence vector into the emotion recognition model to perform customer emotion recognition, wherein the emotion recognition model uses a gated recurrent unit to filter out emotion-irrelevant vectors, and then connects to a normalization layer to output customer emotions.

[0104] In an embodiment of the present invention, the intelligent recognition module 402 is also used to: obtain discrete features for representing customer intentions based on customer sentences; input the discrete features and customer sentence vectors into the intention recognition model to recognize the customer's intentions, wherein the intention recognition model recognizes the customer's intentions in the following manner: for discrete features, high-dimensional mapping is used and then vector dot multiplication is used to perform feature crossover to obtain a first input; for customer speech vectors, a convolutional neural network is used for encoding and connected to a fully connected layer, and the fully connected layer is crossovered with the features to obtain a second input; the first input and the second input are concatenated and connected to the classification layer to obtain an intention recognition result.

[0105] In an embodiment of the present invention, the intelligent recognition module 402 is also used to: input the customer service speech vector into the fine-tuning layer to fine-tune the customer service speech vector; connect the fine-tuning result to the logarithmic loss function for binary classification to obtain the customer service speech detection result.

[0106] In an embodiment of the present invention, the speech quality inspection module 403 is also used to: input the emotion recognition results, intention recognition results, customer service speech detection results, and the customer service speech vector corresponding to the single-sentence customer service speech, as well as the customer sentence vector corresponding to the context customer speech of the single-sentence customer service speech, into the single sentence verification model for single sentence verification, wherein the single sentence verification model performs one-hot coding mapping on the emotion recognition results, intention recognition results, and customer service speech detection results to obtain a one-hot code vector, establishes a gating unit with the one-hot code vector, and sets trainable parameters to establish an association, filters the input content, and uses a multi-head attention mechanism to perform multi-angle attention cross-attention on the filtered content for single sentence verification.

[0107] In an embodiment of the present invention, the speech quality inspection module 403 is also used to: input the customer sentence vector and customer service speech vector, emotion recognition results, intention recognition results, and customer service speech detection results corresponding to the complete conversation into the dialogue verification model for dialogue verification, wherein the dialogue verification model performs one-hot coding mapping on the emotion recognition results, intention recognition results, and customer service speech detection results to obtain a one-hot code vector, establishes a gating unit with the one-hot code vector, and sets trainable parameters to establish an association, filters the input content, and uses a multi-head attention mechanism to perform multi-angle attention cross-attention on the filtered content for dialogue verification.

[0108] In an embodiment of the present invention, the speech quality inspection module 403 is also used to: perform weighted averaging of the single sentence verification results and the dialogue verification results to obtain a quality inspection result for performing online customer service speech quality inspection.

[0109] In an embodiment of the present invention, the device 400 also includes a final result module (not shown in the figure), which is used to: identify positive rules and negative rules for customer service speech quality inspection data; obtain customer service operation records and perform one-hot code mapping on the operation records to obtain a one-hot code vector, and concatenate the one-hot code vector with the rule recognition result to obtain a first quality inspection result; merge the first quality inspection result with the quality inspection result, and match different weights to obtain the final quality inspection result.

[0110] Figure 5 An exemplary system architecture 500 is shown to which the method for quality inspection of online customer service speech or the apparatus for quality inspection of online customer service speech according to the embodiments of the present invention may be applied.

[0111] like Figure 5 As shown, system architecture 500 may include terminal devices 501, 502, 503, a network 504 and a server 505. Network 504 is used to provide a medium for communication links between terminal devices 501, 502, 503 and server 505. Network 504 may include various connection types, such as wired, wireless communication links or optical fiber cables, etc.

[0112] Users can use terminal devices 501, 502, 503 to interact with server 505 through network 504 to receive or send messages, etc. Various communication client applications can be installed on terminal devices 501, 502, 503, such as customer service quality inspection tools, voice quality inspection tools, quality inspection applications, etc. (only examples).

[0113] The terminal devices 501 , 502 , and 503 may be various electronic devices having a display screen and supporting web browsing, including but not limited to smart phones, tablet computers, laptop computers, desktop computers, and the like.

[0114] The server 505 may be a server that provides various services, such as a background management server (only an example) that provides support for quality inspection websites browsed by users using the terminal devices 501, 502, and 503. The background management server may analyze and process the received quality inspection data and other data, and feed back the processing results (such as quality inspection results - only an example) to the terminal device.

[0115] It should be noted that the method for online customer service speech quality inspection provided in the embodiment of the present invention is generally executed by the server 505 , and accordingly, the device for online customer service speech quality inspection is generally set in the server 505 .

[0116] It should be understood that Figure 5 The number of terminal devices, networks and servers in the embodiment is only for illustration. Any number of terminal devices, networks and servers may be provided according to implementation requirements.

[0117] Reference below Figure 6 , which shows a schematic diagram of the structure of a computer system 600 of a terminal device or a server suitable for implementing an embodiment of the present invention. Figure 6 The terminal device or server shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present invention.

[0118] like Figure 6 As shown, the computer system 600 includes a central processing unit (CPU) 601, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 602 or a program loaded from a storage part 608 into a random access memory (RAM) 603. In the RAM 603, various programs and data required for the operation of the system 600 are also stored. The CPU 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.

[0119] The following components are connected to the I / O interface 605: an input section 606 including a keyboard, a mouse, etc.; an output section 607 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 608 including a hard disk, etc.; and a communication section 609 including a network interface card such as a LAN card, a modem, etc. The communication section 609 performs communication processing via a network such as the Internet. A drive 610 is also connected to the I / O interface 605 as needed. A removable medium 611, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 610 as needed, so that a computer program read therefrom is installed into the storage section 608 as needed.

[0120] In particular, according to the embodiments disclosed in the present invention, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the embodiments disclosed in the present invention include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from the network through the communication part 609, and / or installed from the removable medium 611. When the computer program is executed by the central processing unit (CPU) 601, the above-mentioned functions defined in the system of the present invention are executed.

[0121] It should be noted that the computer-readable medium shown in the present invention may be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, 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 above. In the present invention, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, device or device. In the present invention, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, which carries a computer-readable program code. This propagated data signal may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, which may send, propagate or transmit a program for use by or in conjunction with an instruction execution system, apparatus or device. The program code contained on the computer-readable medium may be transmitted using any appropriate medium, including but not limited to: wireless, wire, optical cable, RF, etc., or any suitable combination of the above.

[0122] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present invention. In this regard, each box in the flow chart or block diagram can represent a module, a program segment, or a part of a code, and the above-mentioned module, program segment, or a part of a code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram or flow chart, and the combination of the boxes in the block diagram or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0123] The units or modules involved in the embodiments of the present invention may be implemented in software or hardware. The units or modules described may also be provided in a processor. For example, they may be described as: a processor including a data acquisition module, an intelligent recognition module, and a speech quality inspection module. Among them, the names of these units or modules do not constitute limitations on the units or modules themselves in certain circumstances. For example, the data acquisition module may also be described as "a module for acquiring online customer service speech quality inspection data and performing feature extraction to obtain customer sentence vectors and customer service speech vectors."

[0124] As another aspect, the present invention also provides a computer-readable medium, which may be included in the device described in the above embodiment; or it may exist independently without being assembled into the device. The above computer-readable medium carries one or more programs, and when the above one or more programs are executed by a device, the device includes: obtaining online customer service speech quality inspection data, and performing feature extraction to obtain customer sentence vectors and customer service speech vectors; performing customer emotion recognition and intention recognition based on the customer sentence vectors, and performing customer service speech detection based on the customer service speech vectors; performing single sentence verification and dialogue verification based on the emotion recognition results, intention recognition results, and customer service speech detection results, and performing online customer service speech quality inspection based on the single sentence verification results and dialogue verification results.

[0125] According to the technical solution of the embodiment of the present invention, the online customer service speech quality inspection data is obtained, and feature extraction is performed to obtain customer sentence vectors and customer service speech vectors; customer emotion recognition and intention recognition are performed according to the customer sentence vectors, and customer service speech detection is performed according to the customer service speech vectors; single sentence verification and dialogue verification are performed according to the emotion recognition results, intention recognition results and customer service speech detection results, and online customer service speech quality inspection is performed according to the single sentence verification results and dialogue verification results. The technical means can handle multiple speech compliance quality inspections, solve the tedious problems of manual or rule processing, the dual-channel alignment input method solves the problem of insufficient role feedback detection, and enhances the speech compliance quality inspection ability and accuracy. The device for separate verification of single sentences and dialogues can analyze the quality inspection speech compliance from multiple angles, and the diversified customer service quality inspection models divided by roles and types can effectively save quality inspection time and improve quality inspection results.

[0126] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions may occur depending on design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for quality inspection of online customer service speech, characterized in that: include: Obtain online customer service speech quality inspection data and perform feature extraction to obtain customer sentence vectors and customer service speech vectors; Performing customer emotion recognition and intention recognition based on the customer sentence vector, inputting the customer service speech vector into a fine-tuning layer, and fine-tuning the customer service speech vector; The fine-tuning results are connected to the logarithmic loss function for binary classification to obtain the customer service speech detection results; The emotion recognition result, the intention recognition result, the customer service speech detection result, the customer service speech vector corresponding to the single-sentence customer service speech, and the customer sentence vector corresponding to the context customer speech of the single-sentence customer service speech are input into the single sentence verification model for single sentence verification, wherein the single sentence verification model performs one-hot coding mapping on the emotion recognition result, the intention recognition result, and the customer service speech detection result to obtain a one-hot code vector, establishes a gating unit with the one-hot code vector, and sets trainable parameters to establish association, filters the input content, and uses a multi-head attention mechanism to perform attention cross-talk from multiple angles on the filtered content to perform single sentence verification; The customer sentence vector and customer service speech vector, emotion recognition result, intention recognition result, and customer service speech detection result corresponding to the complete conversation are input into the dialogue verification model for dialogue verification, wherein the dialogue verification model performs one-hot coding mapping on the emotion recognition result, the intention recognition result, and the customer service speech detection result to obtain a one-hot code vector, establishes a gating unit with the one-hot code vector, sets trainable parameters to establish association, filters the input content, and uses a multi-head attention mechanism to perform attention cross-over from multiple angles on the filtered content for dialogue verification; Perform quality inspection on online customer service speech based on single sentence verification results and conversation verification results.

2. The method according to claim 1, characterized in that Before feature extraction, it also includes: Desensitizing the quality inspection data and replacing key information with characteristic characters; Filter non-text content and robot auto-reply messages; Divide the conversation according to the customer's consultation content.

3. The method according to claim 1, characterized in that: Before feature extraction, it also includes: According to the conversation subject, the online customer service speech quality inspection data is divided into customer statements and customer service speech; Feature extraction to obtain customer sentence vectors and customer service speech vectors includes: The pre-trained models are used to extract features of the customer sentences and the customer service speech respectively to obtain customer sentence vectors and customer service speech vectors.

4. The method according to claim 3, characterized in that After extracting features of the customer service speech using the pre-trained model to obtain a customer service speech vector, the method further includes: For the customer service speech vector, the cosine similarity method between vectors is used to evaluate the pre-training effect and adjust the parameters of the pre-training model.

5. The method according to claim 1, characterized in that Before performing emotion recognition and intention recognition of the customer according to the customer sentence vector and performing customer service speech detection according to the customer service speech vector, the method further includes: Annotating the quality inspection data with offensive language; The customer sentence vector and the customer service speech vector are marked with roles and positions.

6. The method according to claim 1, characterized in that Recognizing the customer's emotion based on the customer's sentence vector includes: The customer sentence vector is bidirectionally input into the emotion recognition model to perform customer emotion recognition, wherein the emotion recognition model uses a gated recurrent unit to filter out emotion-irrelevant vectors, and then accesses a normalization layer to output customer emotions.

7. The method according to claim 1, characterized in that Recognizing the customer's intention according to the customer sentence vector includes: Obtain discrete features for representing customer intentions based on customer sentences; The discrete features and the customer sentence vector are input into the intention recognition model to recognize the customer's intention, wherein the intention recognition model recognizes the customer's intention in the following manner: For the discrete features, high-dimensional mapping is performed and then vector dot multiplication is used to perform feature crossover to obtain a first input; For the customer sentence vector, a convolutional neural network is used to encode it and access the fully connected layer, and the fully connected layer is crossed with the feature to obtain a second input; The first input and the second input are concatenated and connected to a classification layer to obtain an intent recognition result.

8. The method according to claim 1, characterized in that The online customer service script quality inspection based on the single sentence verification results and the dialogue verification results includes: The single sentence verification result and the conversation verification result are weighted averaged to obtain a quality inspection result for online customer service speech quality inspection.

9. The method according to claim 8, characterized in that The method further comprises: Identify positive rules and negative rules for the customer service speech quality inspection data; Acquire customer service operation records and perform one-hot code mapping on the operation records to obtain a one-hot code vector, and concatenate the one-hot code vector with the rule recognition result to obtain a first quality inspection result; The first quality inspection result is combined with the quality inspection result, and different weights are matched to obtain a final quality inspection result.

10. A device for quality inspection of online customer service speech, characterized in that: include: The data acquisition module is used to obtain the online customer service speech quality inspection data and perform feature extraction to obtain the customer sentence vector and the customer service speech vector; An intelligent recognition module, used to perform customer emotion recognition and intention recognition based on the customer sentence vector, input the customer service speech vector into a fine-tuning layer, and fine-tune the customer service speech vector; The fine-tuning results are connected to the logarithmic loss function for binary classification to obtain the customer service speech detection results; The speech quality inspection module is used to input the emotion recognition result, the intention recognition result, the customer service speech detection result and the customer service speech vector corresponding to the single-sentence customer service speech, and the customer sentence vector corresponding to the context customer speech of the single-sentence customer service speech into the single-sentence verification model for single-sentence verification, wherein the single-sentence verification model performs one-hot coding mapping on the emotion recognition result, the intention recognition result, and the customer service speech detection result to obtain a one-hot code vector, establishes a gating unit with the one-hot code vector, and sets trainable parameters to establish association, filters the input content, and uses a multi-head attention mechanism to perform attention cross-talk from multiple angles on the filtered content to perform Single sentence verification; the customer sentence vector and customer service speech vector, emotion recognition results, intent recognition results, and customer service speech detection results corresponding to the complete conversation are input into the dialogue verification model for dialogue verification, wherein the dialogue verification model performs one-hot coding mapping on the emotion recognition results, the intent recognition results, and the customer service speech detection results to obtain a one-hot code vector, establishes a gating unit with the one-hot code vector, and sets trainable parameters to establish an association with the customer sentence vector and the customer service speech vector, filters the input content, and uses a multi-head attention mechanism to perform multi-angle attention cross-attention on the filtered content for dialogue verification; conducts online customer service speech quality inspection based on the single sentence verification results and the dialogue verification results.

11. An electronic device for quality inspection of online customer service speech, characterized in that: include: one or more processors; a storage device for storing one or more programs, When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 9.

12. A computer readable medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 9 is implemented.

Images