RPA customer service interaction optimization method and cloud server based on artificial intelligence

Abstract

The present application provides an RPA customer service interaction optimization method and a cloud server based on artificial intelligence. The customer service interaction optimization process is assisted by a sentiment polarity analysis network. When training the neural network, the number of samples required is reduced, and too many labeled customer service text samples are not required, so that the training cost is reduced. Based on some labeled customer service text samples, partial guided clustering is used to obtain the predicted prior information of unlabeled customer service text samples. In this way, the network is trained with the help of massive unlabeled customer service text samples, and the semantic uniformity measurement of the samples as a whole and the paragraphs is considered. The noise of the predicted prior information can be cleaned up, making the predicted prior information more accurate. The trained network has higher accuracy, which helps to obtain accurate sentiment polarity labels for RPA customer service conversation texts, so as to accurately optimize customer service interactions and improve customers' customer service experience.

Description

Technical Field

[0001] The present application relates to but is not limited to the field of data processing technology, and in particular to an artificial intelligence-based RPA customer service interaction optimization method and a cloud server. Background Art

[0002] With the rapid development of information technology, the application of artificial intelligence technology in the field of customer service is becoming more and more extensive, especially the emergence of robotic process automation (RPA) customer service system, which provides enterprises with efficient and convenient customer service solutions. However, in the process of RPA customer service interaction, accurately understanding the emotional polarity of customers is still a very challenging problem.

[0003] When existing sentiment polarity analysis technology is applied to RPA customer service, a large amount of labeled data is usually required to train the neural network model. Obtaining a large number of labeled customer service text samples often requires a lot of manpower, material resources and time costs. For example, manually labeling the sentiment polarity labels of customer service texts not only requires professional labelers, but the labeling process is cumbersome and prone to errors. This has led to the acquisition of training data becoming an important bottleneck restricting the development of sentiment polarity analysis technology in RPA customer service.

[0004] In addition, most traditional sentiment polarity analysis methods rely on large-scale labeled data for supervised learning. When the amount of data is limited, the training effect of the model is difficult to guarantee. Since the performance of the neural network model is highly dependent on the quantity and quality of the training data, limited labeled data may lead to overfitting of the model, that is, the model performs well on the training data, but has poor generalization ability for new data in actual applications. This makes it difficult for existing sentiment analysis technologies to accurately perform sentiment polarity analysis on customer service conversation texts in RPA customer service scenarios, which in turn affects the service quality and customer satisfaction of the RPA customer service system. Summary of the invention

[0005] In view of this, the present application provides an RPA customer service interaction optimization method and cloud server based on artificial intelligence. The technical solution of the present application is implemented as follows:

[0006] On the one hand, the present application provides an artificial intelligence-based RPA customer service interaction optimization method, including: obtaining a basic sentiment polarity analysis network and a customer service text sample library, the customer service text samples in the customer service text sample library include labeled customer service text samples and unlabeled customer service text samples, and the labeled customer service text samples include their corresponding sentiment polarity prior information; for each customer service text sample, performing text splitting on the customer service text sample to obtain multiple text paragraphs of the customer service text sample; using the basic sentiment polarity analysis network, respectively performing text description vector mining on the customer service text sample and each text paragraph of the customer service text sample to obtain a sample description vector of the customer service text sample and a paragraph description vector corresponding to each text paragraph; based on the sentiment polarity prior information of the labeled customer service text sample, performing feature clustering operation on the sample description vector of each customer service text sample in the customer service text sample library to obtain the unlabeled customer service text sample. The method comprises the following steps: obtaining a semantic unity measure between the text paragraph and the corresponding customer service text sample, combining the paragraph description vector corresponding to the text paragraph and the sample description vector of the customer service text sample corresponding to the text paragraph for each text paragraph of the customer service text sample; updating the network weights of the basic sentiment polarity analysis network according to the semantic unity measure and the sentiment polarity prior information corresponding to the unlabeled customer service text sample to obtain a convergent sentiment polarity analysis network; obtaining an RPA customer service conversation text; calling the convergent sentiment polarity analysis network based on the RPA customer service conversation text, and performing sentiment polarity analysis on the RPA customer service conversation text based on the convergent sentiment polarity analysis network to obtain a sentiment polarity label of the RPA customer service conversation text; optimizing the RPA response content based on the sentiment polarity label of the RPA customer service conversation text and a preset customer service reply content mapping strategy.

[0007] On the other hand, the present application provides a cloud server, including a memory and a processor, wherein the memory stores a computer program that can be executed on the processor, and the processor implements the steps in the above method when executing the program.

[0008] When training a neural network, the present application reduces the requirement for the number of samples, and does not require too many labeled customer service text samples, thereby reducing the training cost. Based on some labeled customer service text samples, partial guided clustering is used to obtain predictive prior information for unlabeled customer service text samples. In this way, the network is trained with the help of massive unlabeled customer service text samples, and the semantic uniformity measurement of the samples as a whole and the paragraphs is considered. This can clean up the noise of the predictive prior information, make the predictive prior information more accurate, and increase the accuracy of the trained network, helping to obtain accurate sentiment polarity labels for RPA customer service conversation texts. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] The drawings herein are incorporated into the specification and constitute a part of the specification. These drawings illustrate embodiments consistent with the present application and are used together with the specification to illustrate the technical solution of the present application.

[0010] Figure 1 A schematic diagram of the implementation process of an artificial intelligence-based RPA customer service interaction optimization method provided in an embodiment of the present application.

[0011] Figure 2 A hardware entity diagram of a cloud server provided in an embodiment of the present application. DETAILED DESCRIPTION

[0012] The embodiment of the present application provides an RPA customer service interaction optimization method based on artificial intelligence, which can be executed by a processor of a cloud server. The cloud server can refer to a server, a laptop, a tablet computer, a desktop computer, or other device with data processing capabilities. Figure 1 As shown, the method includes:

[0013] Step S10: Obtain a basic sentiment polarity analysis network and a customer service text sample library. The customer service text samples in the customer service text sample library include labeled customer service text samples and unlabeled customer service text samples. The labeled customer service text samples include their corresponding sentiment polarity prior information.

[0014] The basic sentiment polarity analysis network is a structure specifically used to analyze the sentiment polarity of text. It can be a pre-trained but unconverged neural network, such as a long short-term memory network (LSTM). In the customer service interaction scenario, for the input customer service text, LSTM can process the text information word by word or word by word to capture the semantic information and sentiment tendency in the text. This network contains a series of parameters such as weights and biases. These parameters are based on pre-training in the initial state. For example, in the pre-training process, a large amount of public comment data may be used, which has been labeled as positive or negative sentiment. By training on this data, the network learns the relationship between different vocabulary combinations and sentiment polarity, thereby forming the initial weights and biases. The input to the basic sentiment polarity analysis network is the customer service text.

[0015] The output of the basic sentiment polarity analysis network is the sentiment polarity analysis result of the text. This result may be a numerical value or a classification label. If it is a numerical value, it may be a value between 1 and 1, 1 represents strong negative sentiment, 0 represents neutrality, and 1 represents strong positive sentiment. If it is a classification label, it may be "positive", "negative" or "neutral".

[0016] The customer service text sample library contains labeled customer service text samples and unlabeled customer service text samples. Labeled customer service text samples are samples that have been labeled with sentiment polarity prior information. For example, "This service is very good, I am very satisfied", and its corresponding sentiment polarity prior information is "positive". Another example is "The function of this product is so poor that it cannot be used at all", and the corresponding sentiment polarity prior information is "negative". These labels are obtained through manual labeling or other reliable labeling methods. Unlabeled customer service text samples do not have such clear sentiment polarity markings.

[0017] The cloud server can obtain customer service text samples from multiple channels. On the one hand, they come from actual RPA customer service interaction records. For example, the RPA customer service system of an e-commerce company handles a large number of customer inquiries and feedback every day, and these interaction records can be used as the source of customer service text samples. On the other hand, they can also be obtained from other similar customer service data sets. For example, the customer service data disclosed by other companies in the same industry (in compliance with data use regulations). The customer service text sample library provides a data foundation for the entire RPA customer service interaction optimization. By analyzing the labeled samples in the sample library, the cloud server can learn the relationship between sentiment polarity and text features. For example, certain specific words (such as "satisfied" and "like") are often associated with positive emotions, while words such as "bad" and "unsatisfied" are associated with negative emotions.

[0018] Unlabeled samples provide more data resources for further optimizing the sentiment polarity analysis network. The cloud server can analyze and label unlabeled samples based on the knowledge learned from labeled samples, thereby expanding the number of labeled samples in the sample library and improving the accuracy of sentiment polarity analysis. For example, as unlabeled samples are continuously labeled, when encountering new customer service texts, the cloud server can more accurately determine their sentiment polarity, thereby optimizing the response content of RPA customer service.

[0019] Step S20: For each customer service text sample, the customer service text sample is segmented to obtain multiple text paragraphs of the customer service text sample.

[0020] In RPA customer service scenarios, customer service text samples are often complex and diverse. These texts may contain multiple different semantic parts. Splitting them into multiple text paragraphs helps to analyze the text content more carefully. For example, a customer service text sample may be "I bought your product before. I thought it was good when I first used it. But after using it for a while, I found that one function was not easy to use. I hope you can improve it. In addition, your customer service attitude is still very good." This text contains multiple semantic parts, such as the initial experience of using the product, feedback on problems after using it for a period of time, and evaluation of the customer service attitude. The cloud server can perform more accurate analysis on different semantic parts by splitting such customer service text samples.

[0021] The cloud server can first split the text based on punctuation marks. Common punctuation marks such as periods, exclamation marks, and question marks often mark the end of a semantic paragraph. For example, for the customer service text "This product looks very beautiful. But its price is a bit high, can you give me a discount?" The cloud server can split it into three text paragraphs based on periods and question marks: "This product looks very beautiful.", "But its price is a bit high", "Can you give me a discount?" This splitting method is simple and direct, and can quickly divide the text into basic semantic units.

[0022] The cloud server can also use a splitting method based on semantic understanding, which may involve some models in natural language processing, such as models based on the Transformer architecture. Taking the BERT (Bidirectional Encoder Representations from Transformers) model as an example, the cloud server can use the pre-trained BERT model to encode customer service text samples. The BERT model has learned a lot of language knowledge during the pre-training process and can capture the semantic relationships in the text very well.

[0023] The cloud server inputs the customer service text into the BERT model, and the BERT model outputs the vector representation (word vector) of each word. For example, for the customer service text "I think this product is good, but the delivery speed is a bit slow", the BERT model will generate a corresponding word vector for each word in it. Then, the cloud server can determine the division of text paragraphs by analyzing the semantic relationship between these word vectors. If it is found that the word vector of a certain part is semantically different from other parts, it can be divided into a separate text paragraph. For example, "I think this product is good" and "The delivery speed is a bit slow" have a certain degree of independence in semantics, and the cloud server can split it into two text paragraphs.

[0024] In the RPA customer service interaction scenario, domain-specific vocabulary can also be used as a basis for segmentation. For example, in the customer service interaction of electronic products, some professional vocabulary such as "battery life" and "screen resolution" are used. If the customer service text sample is "I like the appearance design of this phone very much, but the battery life is not ideal, and the screen resolution is not very high", the cloud server can split the text into three text paragraphs based on the domain-specific vocabulary "battery life" and "screen resolution": "I like the appearance design of this phone very much", "but the battery life is not ideal", and "the screen resolution is not very high".

[0025] In subsequent steps, such as the feature clustering operation in step S30, the split text paragraphs can provide richer feature information. Each text paragraph has its own unique semantic features, and the cloud server can perform feature extraction and vector mining for each paragraph.

[0026] Step S30: using a basic sentiment polarity analysis network, respectively mining text description vectors for the customer service text sample and each text paragraph of the customer service text sample, to obtain a sample description vector of the customer service text sample and a paragraph description vector corresponding to each text paragraph.

[0027] The cloud server uses the basic sentiment polarity analysis network obtained in step S10 to mine text description vectors. This basic sentiment polarity analysis network can be, for example, a convolutional neural network (CNN). CNN can effectively capture local features in text in natural language processing. In the customer service text processing scenario, it can identify specific vocabulary combinations and structural features related to sentiment polarity.

[0028] When the cloud server inputs the customer service text sample and its various text paragraphs into the CNN network, the network will perform a series of convolution and pooling operations on the input text. Taking a customer service text sample "The functions of this product are very practical, and I am very satisfied" as an example, CNN will first convert the text into a vector representation. This process may involve mapping each word to a word vector (for example, using a pre-trained word vector model such as Word2Vec). Assume that the word vector corresponding to "this" is [0.1, 0.2, 0.3], and the word vector corresponding to "product" is [0.4, 0.5, 0.6], etc. Then the convolution operation is performed by sliding the convolution kernel of the convolution layer on the matrix composed of these word vectors. Assuming that the size of the convolution kernel is 3x3 (the 3 here represents the dimension of the word vector, and for the sake of simplicity, the dimension is assumed to be 3), the weight matrix of the convolution kernel is , sliding the convolution result on the word vector matrix corresponding to the text. Through the convolution operation, CNN can extract local features in the text, such as the semantic features represented by certain word combinations. After processing through multiple convolutional layers and pooling layers, the sample description vector of the customer service text sample is finally obtained.

[0029] For a text paragraph in the customer service text sample, such as "This product's functions are very practical", the same operation process will obtain the paragraph description vector corresponding to this text paragraph. This paragraph description vector can reflect the semantic features and sentiment-related information of the paragraph. The sample description vector is a comprehensive semantic representation of the entire customer service text sample. The paragraph description vector is a more detailed semantic representation of each text paragraph. Accurate sample description vectors and paragraph description vectors help the cloud server better understand customer needs. In the RPA customer service scenario, when the cloud server receives the customer service text sample of the customer and obtains the corresponding vector representation, it can more accurately grasp the customer's views in various aspects. For example, if a customer's customer service text sample about electronic products is processed and the sample description vector obtained shows that the customer is less satisfied with the product function, and the paragraph description vector further indicates that it is a problem with a specific function, then the RPA customer service can give a targeted response to this problem, such as providing a solution to the function or explaining how to use the function.

[0030] These vectors can also provide a basis for optimizing the content of RPA customer service responses. If vector analysis finds that the customer's emotional polarity on a certain aspect is negative, the RPA customer service response can focus on resolving the customer's dissatisfaction. On the contrary, if it is positive emotional polarity, you can express gratitude or further recommend related products or services. For example, for a customer who is satisfied with the hotel service overall but has doubts about the price, based on the sample description vector and paragraph description vector, the RPA customer service can both affirm the customer's recognition of the service in the response and provide some preferential information on the price or explain the price composition.

[0031] Step S40: Based on the sentiment polarity prior information of the labeled customer service text samples, a feature clustering operation is performed on the sample description vector of each customer service text sample in the customer service text sample library to obtain the sentiment polarity prior information corresponding to the unlabeled customer service text samples.

[0032] The cloud server has already obtained the sample description vectors in the customer service text sample library in the previous step (through step S30), and the labeled customer service text samples have sentiment polarity prior information. The purpose of the feature clustering operation is to use the information of the labeled samples to classify (cluster) the sample description vectors in the entire customer service text sample library, thereby inferring the sentiment polarity prior information of the unlabeled customer service text samples. This helps to expand the number of samples with known sentiment polarity, improve the accuracy of sentiment polarity analysis, and thus optimize RPA customer service interactions.

[0033] The cloud server first deploys several feature cluster centers. These feature cluster centers can be regarded as some representative vectors, which are used to attract and classify other sample description vectors. For example, in RPA customer service interaction, assuming that the customer service text is about hotel services, the cloud server may initially set up three feature cluster centers, representing "positive hotel service experience", "negative hotel service experience" and "neutral hotel service experience". Each feature cluster center is a vector. For example, the feature cluster center vector of "positive hotel service experience" may be [0.8, 0.9, 0.8, 0.7] (the dimensions here represent room satisfaction, service staff attitude satisfaction, catering satisfaction and facility satisfaction, and the numerical value indicates a higher satisfaction).

[0034] Then, the cloud server obtains the matching score between the sample description vector of each customer service text sample in the customer service text sample library and the center of each feature cluster. One way to calculate the matching score can be to calculate the inverse of the Euclidean distance between vectors. Assume that the sample description vector of a customer service text sample is [0.7, 0.8, 0.7, 0.6], and the Euclidean distance formula with the center vector of the "positive hotel service experience" feature cluster [0.8, 0.9, 0.8, 0.7] is , the matching score is s=1 / d.

[0035] For each customer service text sample, the target feature cluster center that matches the customer service text sample is determined based on the matching score and the sentiment polarity prior information of the labeled customer service text sample. For example, if the sentiment polarity prior information of a labeled customer service text sample is "positive", and its sample description vector has the highest matching score with the feature cluster center of "positive hotel service experience", then the cloud server will add this customer service text sample to the customer service text library corresponding to "positive hotel service experience".

[0036] For the customer service text library corresponding to each feature cluster center, the cloud server selects the customer service text samples that meet the feature cluster center conditions in the customer service text library and determines it as the new feature cluster center. For example, in the customer service text library corresponding to "Positive Hotel Service Experience", the cloud server may select those samples whose sample description vectors are closest to the current feature cluster center vector and recalculate a new feature cluster center vector. Assuming that several sample description vectors [0.75, 0.85, 0.78, 0.65], [0.82, 0.88, 0.81, 0.72], etc. are selected from the library, the new feature cluster center vector can be obtained by finding the average value of these vectors, such as .

[0037] The cloud server repeats the process of obtaining matching scores. It stops when the acquired feature cluster center meets the clustering cutoff condition. The clustering cutoff condition can be that the change of the feature cluster center is less than a certain threshold, for example, the Euclidean distance of the feature cluster center vector in two consecutive iterations is less than 0.01. Based on the sentiment polarity prior information of the labeled customer service text samples in the target customer service text library corresponding to the feature cluster center that meets the clustering cutoff condition, the sentiment polarity prior information of the unlabeled customer service text samples in the target customer service text library is determined. For example, if there are many labeled "positive" samples in the customer service text library corresponding to the "positive hotel service experience" feature cluster center that meets the cutoff condition, and an unlabeled sample is also divided into this library, then the cloud server can infer that the sentiment polarity prior information of this unlabeled sample is "positive".

[0038] In the implementation of the second feature clustering operation of the present application, the cloud server determines the sample matching score of each two customer service text samples based on the sample description vectors of each two customer service text samples. For example, for two customer service text samples about hotel services, the sample description vectors are [0.6, 0.7, 0.6, 0.5] and [0.7, 0.8, 0.7, 0.6] respectively, and the sample matching score can be obtained by calculating the cosine similarity of the vectors. The cosine similarity formula is ,in , , the calculated cosine similarity value is the sample matching score of the two samples.

[0039] The cloud server combines the sample matching scores to establish a text topology structure. This text topology structure represents the relationship between each customer service text sample. For example, each customer service text sample can be regarded as a text topology unit, and the lines (line segments) between samples represent their sample matching scores. If the sample matching scores of two samples are high, then the line between them is relatively "thick" (indicating a close relationship), and vice versa. It is relatively "thin". In this hotel service example, customer service text samples with similar sample description vectors will be closer in the text topology structure, and the lines between them represent their degree of similarity. The cloud server combines the text topology structure to conduct prior information transmission on the sentiment polarity prior information of the labeled customer service text samples, and obtains the sentiment polarity prior information corresponding to the unlabeled customer service text samples.

[0040] Specifically, in step S4031, the cloud server uses the line segments between the text topology units in the text topology structure to conduct prior information on the sentiment polarity prior information of the labeled customer service text sample to determine the candidate sentiment polarity prior information of the unlabeled customer service text sample. For example, if an unlabeled sample is closely connected with multiple labeled "positive" samples in the text topology structure (with thicker line segments indicating high sample matching scores), then the candidate sentiment polarity prior information of this unlabeled sample may be "positive".

[0041] For each unlabeled customer service text sample, the candidate sentiment polarity prior information of the unlabeled customer service text sample is adjusted cyclically in combination with the text topology units adjacent to the unlabeled customer service text sample in the text topology structure until the prior information of the text topology units in the text topology structure no longer changes (i.e. converges), and the sentiment polarity prior information corresponding to the unlabeled customer service text sample is obtained. For example, if the initial candidate sentiment polarity prior information of an unlabeled sample is "positive", but some of its adjacent samples become "negative" after information transmission, then the cloud server will readjust the sentiment polarity prior information of this unlabeled sample based on the information of the adjacent samples, and after multiple cyclic adjustments, the sentiment polarity prior information in the entire text topology structure is stabilized. Through feature clustering operations, the cloud server can use the information of labeled samples to infer the sentiment polarity prior information of unlabeled samples. In RPA customer service interactions, this means that more customer service text samples can be accurately labeled with sentiment polarity. For example, for new customer service text samples related to hotel services, even without manual annotation, the cloud server can more accurately judge their sentiment polarity based on the existing clustering results and information of labeled samples, thereby providing a more accurate basis for subsequent response optimization.

[0042] When the cloud server can more accurately determine the sentiment polarity of the customer service text sample, the response content of the RPA customer service can be optimized. If an unlabeled customer service text sample is judged to have negative sentiment polarity (obtained through feature clustering), the RPA customer service can provide solutions or apologies for the customer's dissatisfaction. On the contrary, if it is positive sentiment polarity, it will express gratitude or provide more relevant quality service recommendations. The cloud server performs the feature clustering operation in step S40, which is a key step in the entire RPA customer service interaction optimization process based on artificial intelligence. It helps to improve the accuracy of sentiment polarity analysis, thereby optimizing the effect of RPA customer service interaction.

[0043] Step S50: For each text paragraph of the customer service text sample, the semantic unity measure between the text paragraph and its corresponding customer service text sample is obtained by combining the paragraph description vector corresponding to the text paragraph and the sample description vector of the customer service text sample corresponding to the text paragraph.

[0044] In the previous step, the cloud server has obtained the sample description vector of the customer service text sample (through step S30) and the paragraph description vector corresponding to each text paragraph. These vectors are the basis for measuring semantic unity. For example, in a customer service text sample about electronic products, suppose the entire customer service text sample is "This mobile phone has a beautiful appearance, but the battery life is not very good, and the system sometimes freezes. However, the customer service attitude is not bad, and I hope these problems can be solved." The sample description vector of this sample may be [0.6, 0.3, 0.2, 0.8] (here it is assumed that the four dimensions represent the quantitative evaluation of appearance, battery life, system stability and customer service attitude). For one of the text paragraphs "This mobile phone has a beautiful appearance", its paragraph description vector may be [0.8, 0, 0, 0]. This vector focuses on the characteristics of the appearance, while the weights of other aspects are 0, because the paragraph mainly describes the appearance.

[0045] The cloud server establishes a sample description vector library based on the sample description vectors of each customer service text sample, and establishes a paragraph description vector library based on the paragraph description vectors of the text paragraphs of each customer service text sample. In the above example of the electronic product customer service text sample, if there are multiple such customer service text samples, the cloud server will store the sample description vector of each sample in the sample description vector library, and store the paragraph description vector of each text paragraph in the paragraph description vector library.

[0046] For each customer service text sample, the cloud server indexes at least one target sample description vector similar to the sample description vector of the customer service text sample in the sample description vector library, and determines the pending sentiment polarity prior information set of the customer service text sample based on the sentiment polarity prior information of the customer service text sample corresponding to the target sample description vector. For each text paragraph of the customer service text sample, the cloud server indexes at least one target paragraph description vector similar to the paragraph description vector of the text paragraph in the paragraph description vector library, and determines the pending sentiment polarity prior information set of the text paragraph based on the sentiment polarity prior information of the customer service text sample corresponding to the target paragraph description vector.

[0047] The cloud server obtains a semantic unity measure between a text paragraph and its corresponding customer service text sample based on the pending sentiment polarity prior information set of the text paragraph and the pending sentiment polarity prior information set of the customer service text sample corresponding to the text paragraph.

[0048] First, the cloud server determines the number of occurrences of the sentiment polarity prior information in the pending sentiment polarity prior information set of the text paragraph, and determines the sentiment polarity probability distribution of the text paragraph. In the above example, the pending sentiment polarity prior information set of the text paragraph "This mobile phone has a beautiful appearance" is {"positive"}, so its sentiment polarity probability distribution is that the positive probability is 1 and the negative probability is 0. Then, the number of occurrences of the sentiment polarity prior information in the pending sentiment polarity prior information set of the customer service text sample corresponding to the text paragraph is determined, and the sentiment polarity probability distribution of the customer service text sample corresponding to the text paragraph is determined. The pending sentiment polarity prior information set of the customer service text sample is {"negative", "mixed"}. Assume that after statistics, the probability of negative occurrence is 0.6 and the probability of mixed occurrence is 0.4. Finally, based on the sentiment polarity probability distribution of the text paragraph and the sentiment polarity probability distribution of the customer service text sample corresponding to the text paragraph, the semantic unity measure between the text paragraph and its corresponding customer service text sample is obtained. A simple calculation method can be to calculate the Euclidean distance of the probability distributions of the two. Suppose the probability distribution of the text paragraph is (here =1, =0)), the probability distribution of the customer service text sample is (here =0.6, =0.4), the Euclidean distance formula is ,here This distance value can be used as a measure of semantic unity, and the smaller the distance, the higher the semantic unity.

[0049] The semantic unity measure provides a more comprehensive perspective for sentiment polarity analysis. In the example of the electronic product customer service text above, if the sentiment polarity is analyzed based solely on the paragraph description vector of the text paragraph "This phone looks beautiful", it may be simply judged as positive. However, by calculating the semantic unity measure with the entire customer service text sample, it is found that there is a certain difference in the probability distribution of sentiment polarity with the entire sample (such as the Euclidean distance calculated above), which means that although the paragraph itself looks positive, it may be affected by other negative factors in the context of the entire customer service text sample. This comprehensive consideration helps to more accurately analyze the sentiment polarity of the entire customer service text sample and avoid one-sided judgment based on a certain paragraph.

[0050] Semantic uniformity measurement also plays an important role in optimizing the content of RPA customer service responses. If the semantic uniformity measurement between a text paragraph and its corresponding customer service text sample is low, it means that the paragraph may have a large semantic deviation from the whole. The cloud server performs the semantic uniformity measurement operation in step S50, which is an important part of the RPA customer service interaction optimization process based on artificial intelligence. Through this measurement, the cloud server can have a deeper understanding of the semantic relationship between each text paragraph within the customer service text sample and the entire sample, which helps to improve the accuracy of sentiment polarity analysis, thereby optimizing the response content of RPA customer service and improving customer satisfaction.

[0051] Step S60: Based on the semantic unity metric and the sentiment polarity prior information corresponding to the unlabeled customer service text samples, the network weights of the basic sentiment polarity analysis network are updated to obtain a converged sentiment polarity analysis network.

[0052] In the AI-based RPA customer service interaction optimization scenario, the cloud server executing step S60 is a key step, which aims to optimize the basic sentiment polarity analysis network by utilizing the semantic uniformity measurement obtained in the previous step and the sentiment polarity prior information of the unlabeled customer service text samples, thereby obtaining a more accurate converged sentiment polarity analysis network.

[0053] The basic sentiment polarity analysis network in the cloud server (for example, a multi-layer perceptron neural network) has been used in the previous steps to analyze customer service text samples and text paragraphs, and obtained results such as sample description vectors and paragraph description vectors. The initial weights of this network are determined based on pre-training or initialization strategies, but may not be accurate enough and need to be adjusted based on new information. In RPA customer service interactions, the accuracy of the network weights directly affects the accuracy of the sentiment polarity analysis of customer service texts. For example, for a customer service text about hotel services, "The hotel rooms are clean, but the service staff has a bad attitude", if the network weights are inaccurate, the sentiment polarity of the entire text may be incorrectly judged. By updating the network weights, the accuracy of the sentiment polarity judgment of such texts can be improved.

[0054] The cloud server first performs sentiment polarity analysis on the customer service text sample based on the sample description vector of the customer service text sample to obtain the sample sentiment polarity probability distribution corresponding to the customer service text sample. For example, for a hotel customer service text sample "The hotel rooms are spacious and bright, the facilities are complete, and the breakfast is rich in variety, but the price is a bit high", its sample description vector is [0.8, 0.8, 0.7, 0.3] (here it is assumed that the dimensions represent the evaluation quantification of the room, facilities, breakfast and price respectively). The cloud server can obtain the sample sentiment polarity probability distribution through the activation function (such as the Softmax function) in the basic sentiment polarity analysis network. Assume that the Softmax function is ,in is the unnormalized log probability of the network output, and n is the number of sentiment polarity categories (such as positive, negative, neutral, (n = 3)). After calculation, the probability of positive sentiment is 0.6, the probability of negative sentiment is 0.3, and the probability of neutral sentiment is 0.1.

[0055] Similarly, the cloud server performs sentiment polarity analysis on the text paragraphs based on the paragraph description vectors corresponding to the text paragraphs, and obtains the paragraph sentiment polarity probability distribution corresponding to the text paragraphs.

[0056] The cloud server updates the network weights of the basic sentiment polarity analysis network by combining the sample sentiment polarity probability distribution, paragraph sentiment polarity probability distribution, semantic unity measurement, and sentiment polarity prior information corresponding to the unlabeled customer service text samples to obtain a converged sentiment polarity analysis network.

[0057] Specifically, in step S631, for each customer service text sample, the cloud server optimizes the sentiment polarity prior information of the customer service text sample based on the paragraph sentiment polarity probability distribution of each text paragraph in the customer service text sample and the semantic unity metric corresponding to each text paragraph, and obtains the optimized sentiment polarity prior information of the customer service text sample. For example, for the above-mentioned hotel customer service text sample, the paragraph sentiment polarity probability distribution of the paragraph "The hotel's rooms are spacious and bright" is a positive probability of 0.9. Assuming that it is consistent with the semantic unity metric of the entire sample (calculated by the previous step S50, for example, 0.2, indicating that the semantic unity is high), according to certain rules (for example, if the semantic unity metric is less than a certain threshold of 0.3, then a higher weight is given to the paragraph sentiment polarity), the sentiment polarity prior information of the entire sample can be adjusted. If the original sample's sentiment polarity prior information is mixed (both positive and negative), it may be more inclined to be positive after adjustment.

[0058] In step S632, the cloud server integrates the optimized sentiment polarity prior information of each customer service text sample according to the sample sentiment polarity probability distribution corresponding to each customer service text sample to obtain the fusion training cost. For example, assuming there are N customer service text samples, for the i-th customer service text sample, its sample sentiment polarity probability distribution is (corresponding to positive, negative, and neutral probabilities respectively), the optimized sentiment polarity prior information is (For example, Can be a value between 1 and 1, indicating the sentiment polarity tendency, 1 is strongly negative, 1 is strongly positive). One way to calculate the fusion training cost C can be .

[0059] In step S633, the cloud server updates the network weights of the basic sentiment polarity analysis network according to the fusion training cost to obtain a converged sentiment polarity analysis network. For example, in a simple gradient descent algorithm, if the network weight is w and the learning rate is α, the weight update formula is ,in It is the partial derivative of the fusion training cost C with respect to the weight w. The cloud server iterates this process multiple times and continuously adjusts the network weights until certain convergence conditions are met (such as the change in the fusion training cost is less than a certain threshold).

[0060] By updating the network weights, the cloud server can improve the accuracy of the basic sentiment polarity analysis network for sentiment polarity analysis of customer service texts. In RPA customer service interactions, it is crucial to accurately judge the sentiment polarity of customer texts. For example, for an RPA customer service system of an e-commerce platform, when a customer sends a text such as "This product looks beautiful, but the quality is not very good", the updated network can more accurately judge that the overall sentiment polarity is negative, rather than being misled by the positive description of "The appearance is beautiful". This helps RPA customer service make more appropriate responses, such as providing solutions or explanations for quality issues.

[0061] More accurate sentiment polarity analysis can optimize the response strategy of RPA customer service. If the sentiment polarity is judged incorrectly, RPA customer service may give inappropriate responses. For example, if a customer service text with negative sentiment is mistakenly judged as positive, RPA customer service may recommend more products instead of solving the customer's problem. The convergent sentiment polarity analysis network after weight update can avoid this situation, allowing RPA customer service to provide appropriate responses such as solutions, apologies or thanks based on accurate sentiment polarity.

[0062] The cloud server performs the network weight update operation in step S60, which is the core link in the entire RPA customer service interaction optimization process based on artificial intelligence. It continuously optimizes the basic sentiment polarity analysis network by integrating the semantic uniformity measurement and the sentiment polarity prior information of unlabeled samples, thereby improving the accuracy of sentiment polarity analysis and further optimizing the response content and interaction effect of RPA customer service.

[0063] Step S70: Obtain the RPA customer service conversation text.

[0064] In the RPA customer service system, the RPA customer service conversation text mainly comes from the interaction records between customers and customer service robots or automated customer service systems. These texts may be stored in a database, and the cloud server can obtain these text data from the database. In addition, in some enterprise-level customer service systems, customer service interactions from multiple channels may be involved, such as online customer service on websites and mobile APP customer service. Customer interaction texts from different channels will be aggregated into a unified storage system, and the cloud server can obtain all RPA customer service conversation texts from this unified storage system to ensure that customers' needs and emotional tendencies can be fully analyzed.

[0065] The cloud server obtains the RPA customer service conversation text through a specific data interface. This data interface is pre-defined in the RPA customer service system architecture, which ensures that the cloud server can obtain the required data safely and efficiently. For example, in a customer service system based on a microservice architecture, there may be a microservice dedicated to data acquisition. The cloud server obtains the RPA customer service conversation text by calling the interface of this microservice. At the same time, permission management plays an important role in this process. Only cloud servers with appropriate permissions can obtain these conversation texts. Permission management can be based on the role-based access control (RBAC) model. For example, within an enterprise, only a specific cloud server role responsible for customer service data analysis and optimization may be granted permission to obtain RPA customer service conversation texts. This permission management ensures data security and compliance and prevents data leakage and abuse. After obtaining the RPA customer service conversation text, the cloud server may perform some preprocessing and cleaning operations. Since the conversation text may come from different channels and users, the quality and format of the data may vary. For example, there may be some special characters (such as emoticons, garbled characters, etc.) or extra spaces. The cloud server can handle these problems by writing scripts or using data cleaning tools.

[0066] Suppose there is a customer service conversation text containing emoticons, "This product looks good, but the price is a bit high." The cloud server can use regular expressions and other technologies to remove emoticons and process the text into "This product looks good, but the price is a bit high." For redundant spaces, such as "this product," you can write a program to replace multiple consecutive spaces with a single space to get "this product." These preprocessing and cleaning operations help improve the accuracy and efficiency of subsequent analysis.

[0067] Step S80: calling a convergent sentiment polarity analysis network based on the RPA customer service conversation text, and performing sentiment polarity analysis on the RPA customer service conversation text based on the convergent sentiment polarity analysis network to obtain a sentiment polarity label of the RPA customer service conversation text.

[0068] The cloud server has obtained a converged sentiment polarity analysis network in the previous step (step S60). This network is obtained through a series of training and optimization processes, and it has the ability to perform relatively accurate sentiment polarity analysis on customer service texts.

[0069] The cloud server inputs the obtained RPA customer service conversation text into the convergent sentiment polarity analysis network. For example, suppose the RPA customer service conversation text is "This hotel has a good location, but the room is a bit small." This text will be processed according to the network input requirements. It may first be segmented. If a pre-trained word vector model (such as Word2Vec) is used, each word will be mapped to a corresponding word vector. Assume that the word vector corresponding to "this" is [0.1, 0.2,0.3], and the word vector corresponding to "hotel" is [0.4, 0.5, 0.6], etc. (here are just simple examples of word vectors). Then these word vectors will be organized into a sequence according to the order in the text and input into the convergent sentiment polarity analysis network. Inside the convergent sentiment polarity analysis network, for a network structure such as LSTM, it will gradually process the input word vector sequence. The LSTM unit will process information according to its internal gating mechanism (including input gate, forget gate, and output gate). For example, for the content "This hotel has a good location", the network will use the gating mechanism to decide which information needs to be retained, updated, or forgotten. During the processing, the network will continuously update its internal hidden state, which can be regarded as a comprehensive representation of the text content that has been processed previously.

[0070] When processing "but the room is a bit small", the network will continue to adjust the hidden state based on the previous hidden state and the current input word vector information. In this process, the network will capture the semantic changes brought about by the transition word "but" and comprehensively consider it with the previous information about the good location of the hotel.

[0071] After being processed by the network, a sentiment polarity label about the RPA customer service conversation text will be output. In the above hotel example, based on the positive comments about the hotel location and the negative comments about the room size in the text, the network may output a mixed (both positive and negative) sentiment polarity label.

[0072] The sentiment polarity label has a direct guiding significance for the response of RPA customer service. If the sentiment polarity label is positive, such as in the example of "The service of this hotel is very thoughtful, the room is also very clean, I am very satisfied", RPA customer service can express gratitude in the response, and can further recommend other services or preferential activities of the hotel, such as "Thank you for your recognition of our hotel. Our hotel currently has special catering services, you can try it."

[0073] If the sentiment polarity label is negative, for example, for a mixed sentiment polarity situation such as "This hotel has a great location, but the room is a bit small", the RPA customer service can first affirm the customer's recognition of the hotel's location in the response, and then provide a solution or explanation for the problem of the small room, such as "Thank you for your recognition of the location of our hotel. Regarding the problem of small rooms, we have different room types to choose from. Next time, you can try to book our deluxe room type, which will be more spacious and comfortable."

[0074] Step S90: According to the sentiment polarity label of the RPA customer service conversation text, based on the preset customer service reply content mapping strategy, the RPA response content is optimized.

[0075] In the AI-based RPA customer service interaction optimization scenario, the cloud server executing step S90 is the key link in the entire process, which directly affects the quality and effect of the interaction between RPA customer service and customers.

[0076] In step S80, the cloud server has obtained the sentiment polarity label of the RPA customer service conversation text, which reflects the customer's sentiment tendency in the conversation, such as positive, negative or neutral. This sentiment tendency is an important basis for the cloud server to optimize the RPA response content.

[0077] Take the customer service interaction of an e-commerce platform as an example. If the customer's conversation text is "The quality of this product is very good, I am very satisfied, and the delivery speed is also very fast", the sentiment polarity label obtained after analysis is positive. This positive label indicates that the customer has a positive attitude towards the product and service. On the contrary, if the customer says "The function of this product is not very easy to use, it is very different from what I expected", the sentiment polarity label obtained is negative, which means that the customer is dissatisfied.

[0078] A rule-based mapping strategy is a set of predefined rules that associate different sentiment polarity labels with corresponding response content. For example, for a positive sentiment polarity label, the response content can focus on expressing gratitude, providing additional discounts, or recommending related products. Assuming that in an e-commerce scenario, when the sentiment polarity label is positive, the preset response may be "Thank you very much for your recognition of our products and services! To express our gratitude, we provide you with a coupon for your next purchase. In addition, we have some high-quality products related to the product you purchased that may be of interest to you."

[0079] For negative sentiment polarity labels, the response should focus on solving the customer's problem, apologizing, or providing a solution. For example, if the customer is dissatisfied with a product feature, the response may be "We are sorry to give you a bad experience. Our technical team is aware of this problem and is working hard to improve it. In the meantime, we can provide you with some temporary solutions, you can try [specific solution]."

[0080] For neutral sentiment polarity labels, the response content can be to further inquire about the customer's needs or provide more product information. For example, "You mentioned that the product is OK, do you have any other expectations or suggestions? Here are some detailed information about the product that we can provide you."

[0081] Cloud servers can also use machine learning models to build customer service response content mapping strategies, such as decision tree models. Decision tree models can be trained based on a large amount of historical customer service interaction data, which includes sentiment polarity labels and corresponding high-quality response content.

[0082] When training a decision tree model, the input features can be sentiment polarity labels and other possible related features, such as the customer's purchase history, product type, etc. For example, for an e-commerce platform selling electronic products, the input of the decision tree model may include the customer's sentiment polarity label for electronic products, the frequency of customers' purchase of electronic products, whether the customer has purchased high-end products, etc.

[0083] Each node of the decision tree is split according to different feature values, and the final leaf node corresponds to the specific reply content. For example, for a customer who often buys high-end electronic products and gives a negative sentiment polarity label to a product, the decision tree model may recommend a high-end alternative product and provide a certain discount as the reply content.

[0084] When the sentiment polarity label of the RPA customer service conversation text is positive, the cloud server optimizes the RPA response content according to the preset mapping strategy. For example, in the travel service scenario, if the customer says "This trip was very pleasant, the tour guide service was very thoughtful, and the accommodation conditions were also very good", the cloud server optimizes the RPA response content based on the positive sentiment polarity label and the preset rule-based mapping strategy, which may be "I am very glad that you have such a wonderful travel experience! We will continue to work hard to maintain this service quality. If you want to travel again in the future, we can recommend some other popular travel routes for you."

[0085] If we use a mapping strategy based on a decision tree model and take into account the customer's travel destination preference (assuming this is also an input feature of the decision tree model), if the customer prefers seaside travel and the current travel destination is in the mountains, the RPA response may be "Thank you for your satisfaction with this trip! You seem to like traveling very much. We found that you previously preferred seaside travel. We have some great seaside travel routes recommended to you, which may be of interest to you."

[0086] If a mapping strategy based on a decision tree model is used, and the decision tree model takes into account the busy hours of the restaurant (as an input feature), if the service is slow due to the busy hours of the restaurant, the RPA response may be "We are sorry that you have encountered slow service during the busy hours. We will increase staff during the busy hours to improve service efficiency. At the same time, regarding the taste of the dishes, we will make improvements based on your feedback. We hope you can give us a chance to improve, and we will provide you with a free appetizer next time you dine in."

[0087] By optimizing RPA response content based on sentiment polarity labels, customers' needs and emotional expectations can be better met. When customers express positive emotions, positive responses will make them feel valued and recognized; when customers express negative emotions, timely solutions and apologies will ease their dissatisfaction; for neutral emotions, further inquiries and improvement measures demonstrate respect for customer opinions. This helps improve customer satisfaction with customer service and enhance customer loyalty to the company.

[0088] As an implementation mode, step S40, based on the sentiment polarity prior information of the labeled customer service text samples, performs a feature clustering operation on the sample description vectors of each customer service text sample in the customer service text sample library to obtain the sentiment polarity prior information corresponding to the unlabeled customer service text samples, which may specifically include:

[0089] Step S41: deploying a number of feature cluster centers, and obtaining a matching score between a sample description vector of each customer service text sample in a customer service text sample library and each feature cluster center;

[0090] Step S42: for each customer service text sample, according to the matching score and the sentiment polarity prior information of the labeled customer service text sample, determine the target feature cluster center that matches the customer service text sample in each feature cluster center, and add the customer service text sample to the customer service text library corresponding to the target feature cluster center;

[0091] Step S43: for the customer service text library corresponding to each feature cluster center, select a customer service text sample that meets the feature cluster center condition in the customer service text library and determine it as a new feature cluster center;

[0092] Step S44: Repeat the process of obtaining the matching score between the sample description vector of each customer service text sample in the customer service text sample library and each feature cluster center, and stop when the obtained feature cluster center meets the clustering cutoff condition. Determine the sentiment polarity prior information of the unlabeled customer service text samples in the target customer service text library based on the sentiment polarity prior information of the labeled customer service text samples in the target customer service text library corresponding to the feature cluster center that meets the clustering cutoff condition.

[0093] In the RPA customer service interaction scenario, the cloud server first needs to deploy feature cluster centers in order to effectively classify and analyze customer service text samples. These feature cluster centers can be regarded as representative points, which represent different types of customer service text feature sets in the feature space. For example, in the context of e-commerce customer service, if the customer service text mainly involves product evaluation, logistics inquiry, and after-sales service, the cloud server may deploy three feature cluster centers, corresponding to these three main types of customer service interactions.

[0094] Assume that the center of the first feature cluster represents text related to product evaluation. Its feature vector may emphasize lexical features related to product attributes (such as quality, function, etc.). If lexical features are represented by vectors, the feature "product quality" may be represented as [0.8, 0, 0] (here it is assumed that the vector dimensions represent product quality, logistics speed, and after-sales service, respectively, and only the product quality dimension has a high weight). The center of the second feature cluster represents text related to logistics queries. Its feature vector may have a high weight on lexical features related to logistics speed, such as [0, 0.8, 0]. The center of the third feature cluster represents text related to after-sales service. Its feature vector may be [0, 0, 0.8]. The setting of these feature cluster center vectors is based on the understanding of the content of RPA customer service interaction and the abstract representation of different business types.

[0095] The cloud server obtains the matching score between the sample description vector of each customer service text sample in the customer service text sample library and the center of each feature cluster. A common way to calculate the matching score is to use a metric such as Euclidean distance or cosine similarity.

[0096] For each customer service text sample, the cloud server determines the target feature cluster center based on the matching score calculated in step S41 and the sentiment polarity prior information of the labeled customer service text sample. For example, suppose there is a labeled customer service text sample whose sentiment polarity prior information is "positive", and the sample description vector of the sample has the highest matching score with the feature cluster center representing the product evaluation. This indicates that this customer service text sample is related to the product evaluation in terms of content features, and due to its positive sentiment polarity prior information, the cloud server can determine that the feature cluster center related to the product evaluation is the target feature cluster center.

[0097] In another example, the content of a customer service text sample is "I am very satisfied with your after-sales service and the problem was handled very quickly." After the sample description vector is calculated, the matching score with the feature cluster center related to after-sales service is the highest, and the sentiment polarity prior information of this sample is known to be "positive", then the feature cluster center related to after-sales service is the target feature cluster center of this customer service text sample.

[0098] Once the target feature cluster center is determined, the cloud server will add this customer service text sample to the customer service text library corresponding to the target feature cluster center. For example, for the customer service text sample with positive reviews on after-sales service mentioned above, the cloud server will add it to the customer service text library corresponding to the after-sales service feature cluster center. This customer service text library will contain all customer service text samples related to after-sales service and whose target feature cluster center has been determined to be after-sales service. As more customer service text samples are added to the corresponding customer service text libraries, these libraries will gradually accumulate texts with similar features and sentiment polarity, providing a data basis for subsequent analysis and processing.

[0099] In the customer service text library corresponding to each feature cluster center, the cloud server needs to filter out customer service text samples that meet the feature cluster center conditions. This condition can be set according to specific business needs and data analysis goals. For example, in the e-commerce customer service scenario, for the customer service text library corresponding to the product evaluation feature cluster center, the cloud server may set the condition of selecting customer service text samples that have clear and representative evaluations of a key attribute of the product (such as product quality).

[0100] Assume that in the customer service text library corresponding to the center of the product evaluation feature cluster, there is a customer service text sample "The quality of this product is super good, I like it very much", and this sample has a clear and positive evaluation of the product quality. If the cloud server sets the condition of selecting samples that have a positive and clear evaluation of product quality, then this sample is likely to meet the condition of the feature cluster center.

[0101] Customer service text samples that meet the conditions will be determined as the new feature cluster center. Continuing with the above example, if this positive evaluation sample about product quality is selected, it will become part of the new feature cluster center. The new feature cluster center will be re-determined by combining the feature vectors of these samples that meet the conditions. Assuming that several samples that meet the conditions are selected, their sample description vectors are [0.8, 0.1, 0.1], [0.9, 0.05, 0.05] and [0.85, 0.1,0.05] respectively. The new feature cluster center vector can be obtained by calculating the average value of these vectors. This method of re-determining the feature cluster center enables the feature cluster center to better represent the text features in the current customer service text library. With the continuous accumulation and updating of data, the feature cluster center will more accurately reflect the actual distribution of customer service text.

[0102] The cloud server repeats the process of obtaining the matching score between the sample description vector and the feature cluster center. This is because as the number of customer service text samples increases or as the feature cluster center is updated, the previous matching relationship may change. For example, as new customer service text samples are added to each customer service text library, these new samples may change the feature distribution of the entire customer service text library, thereby affecting the matching relationship between the sample and the feature cluster center.

[0103] Assume that during multiple iterations, new customer service text samples are added to the customer service text library corresponding to the product evaluation feature cluster center. These new samples may contain some product attribute evaluations that were not fully considered before. When recalculating the matching score, some samples that originally had a low matching score with the product evaluation feature cluster center may change their matching scores due to the addition of new samples. This repeated calculation can ensure that the matching relationship between the feature cluster center and the customer service text samples always remains accurate.

[0104] The clustering cutoff condition is the criterion for deciding when to stop the repeated calculation of the matching score process. This condition can be set in many ways. For example, it can be set to stop when the change of the feature cluster center is less than a certain threshold. Specifically, the Euclidean distance between the new feature cluster center vector and the feature cluster center vector obtained in the previous iteration is calculated. If this distance is less than a preset threshold (such as 0.01), it is considered that the feature cluster center has stabilized and meets the clustering cutoff condition.

[0105] Another clustering cutoff condition is to stop when the distribution ratio of customer service text samples between the centers of each feature cluster changes less than a certain percentage (such as 5%) in several consecutive iterations (such as 5 times). This means that after multiple iterations, the distribution of customer service text samples between the centers of different feature clusters has become stable and no longer changes significantly.

[0106] When the clustering cutoff condition is met, the cloud server determines the sentiment polarity prior information of the unlabeled customer service text samples in the target customer service text library based on the sentiment polarity prior information of the labeled customer service text samples in the target customer service text library corresponding to the feature cluster center that meets the clustering cutoff condition. For example, in the customer service text library corresponding to the product evaluation feature cluster center, when the clustering cutoff condition is met, it is found that the sentiment polarity prior information of most (e.g., 70%) of the labeled customer service text samples in this customer service text library is "positive". If there are some unlabeled customer service text samples in this customer service text library, the cloud server can infer that the sentiment polarity prior information of these unlabeled customer service text samples is also likely to be "positive".

[0107] In another customer service text library corresponding to the center of the logistics query feature cluster, if the clustering cutoff condition is met, 50% of the labeled customer service text samples are "negative", 30% are "positive", and 20% are "neutral", for the unlabeled customer service text samples, the cloud server can preliminarily determine their sentiment polarity prior information based on this proportion distribution. For example, if an unlabeled customer service text sample is in this customer service text library, it has a greater probability (50%) of being considered to have "negative" sentiment polarity prior information.

[0108] By executing step S40 of this implementation method, the cloud server can effectively use the sentiment polarity prior information of the labeled customer service text samples to perform feature clustering operations on the sample description vectors in the customer service text sample library, thereby obtaining the sentiment polarity prior information of the unlabeled customer service text samples. This is of great significance for subsequent sentiment polarity analysis, RPA customer service response optimization and other links. For example, accurate sentiment polarity prior information of unlabeled samples can improve the accuracy of the entire sentiment polarity analysis, allowing RPA customer service to better provide appropriate response content based on the customer's emotional tendencies, thereby improving customer satisfaction.

[0109] As another implementation, step S40, based on the sentiment polarity prior information of the labeled customer service text samples, performs a feature clustering operation on the sample description vectors of each customer service text sample in the customer service text sample library to obtain the sentiment polarity prior information corresponding to the unlabeled customer service text samples, which may specifically include:

[0110] Step S401: determining a sample matching score between every two customer service text samples based on the sample description vectors between every two customer service text samples;

[0111] Step S402: Building a text topology structure based on sample matching scores, where the text topology structure represents the relationship between each customer service text sample;

[0112] Step S403: Combined with the text topology structure, the sentiment polarity prior information of the labeled customer service text sample is transmitted to obtain the sentiment polarity prior information corresponding to the unlabeled customer service text sample.

[0113] In RPA customer service interaction, the sample description vector of the customer service text sample is a quantitative representation of the customer service text content. These vectors are mined by the basic sentiment polarity analysis network in the previous step (such as step S30). For example, for a customer service text sample about hotel service, "The hotel room is very clean and the service staff's attitude is also very good", its sample description vector may be [0.8, 0.9, 0] (here it is assumed that the three dimensions represent the room condition, the service staff's attitude and other uninvolved aspects, and the numerical value represents the quantification of the satisfaction degree). This vector can summarize the situation of the customer service text sample in different features.

[0114] The cloud server determines the sample matching score between each two customer service text samples based on their sample description vectors. A common method is to calculate the cosine similarity between vectors. Assume that there are two customer service text samples, and the sample description vector of sample A is , the sample description vector of sample B is , then the cosine similarity formula between them is ,in The value of this cosine similarity is the sample matching score between sample A and sample B.

[0115] In step S402, the cloud server establishes a text topology structure in combination with the sample matching score. In this structure, each customer service text sample can be regarded as a node (i.e., a text topology unit), and the sample matching score between samples determines the nature of the line (i.e., line segment) between the nodes. For example, if the sample matching score of two customer service text samples is high, then the line between them can be represented as thicker or shorter (in the visual representation), indicating that the relationship between them is close; conversely, if the sample matching score is low, the line is thinner or longer, indicating that the relationship is distant.

[0116] Assume that in an e-commerce customer service scenario, there are customer service text samples A, B, C, D, etc. The sample matching score between sample A and sample B is high, and the sample matching score between sample A and sample C is low. In the constructed text topology, the connection between node A and node B will be "tighter" than the connection between node A and node C. This text topology can intuitively show the relationship network between each customer service text sample.

[0117] This text topology provides a basis for the subsequent transmission of sentiment polarity prior information. It enables the cloud server to grasp the association pattern between customer service text samples as a whole. For example, if some customer service text samples are found to form a close substructure (subgraph) in the text topology, this may mean that these samples are similar in semantics and sentiment. By analyzing this structure, the cloud server can better understand the distribution pattern of customer service text samples, thereby more effectively transmitting sentiment polarity prior information.

[0118] In step S403, the sentiment polarity prior information of the labeled customer service text sample is transmitted in combination with the text topology structure to obtain the sentiment polarity prior information corresponding to the unlabeled customer service text sample.

[0119] In the text topology, the cloud server uses the connection (line segment) between nodes to transmit the sentiment polarity prior information of the labeled customer service text samples. For example, suppose that in a customer service text sample library about travel services, the sentiment polarity prior information of the labeled customer service text sample A is "positive", and it has a connection with an unlabeled customer service text sample B in the text topology, and this connection indicates that there is a high sample matching score between them (i.e., a close relationship). Then, the unlabeled customer service text sample B will be affected by the positive sentiment polarity prior information of sample A, and the candidate sentiment polarity prior information of B may be preliminarily determined to be "positive". For another example, the sentiment polarity prior information of the labeled customer service text sample C is "negative", and the connection between it and the unlabeled customer service text sample D indicates that although the sample matching score between them is not very high, there is also a certain correlation. In this case, the unlabeled customer service text sample D will also be affected by the negative sentiment polarity prior information of C to a certain extent, and its candidate sentiment polarity prior information may be inclined to the "negative" direction, but it may not be as strong as when it is associated with a high matching score sample.

[0120] For each unlabeled customer service text sample, the cloud server will cyclically adjust its candidate sentiment polarity prior information in combination with its adjacent text topology units in the text topology structure. For example, the unlabeled customer service text sample E has three adjacent text topology units in the text topology structure, namely the labeled customer service text sample F (sentiment polarity prior information is "positive"), G (sentiment polarity prior information is "negative") and the unlabeled customer service text sample H (whose candidate sentiment polarity prior information was preliminarily determined to be "positive" in the previous transmission).

[0121] In the first round of adjustment, assume that the candidate sentiment polarity prior information of sample E is adjusted based on the sample matching scores of sample E and the three adjacent units and their sentiment polarity prior information or candidate sentiment polarity prior information. If sample E has a high sample matching score with sample F, a low sample matching score with sample G, and a medium sample matching score with sample H, then the candidate sentiment polarity prior information of sample E may be more inclined to "positive", but due to the association with sample G, the "negative" effect will not be completely ignored. Then the cloud server will continue to repeat this adjustment process, and each adjustment will take into account the relationship between the unlabeled customer service text sample and all adjacent units. This process will continue until the prior information of all text topology units in the text topology structure no longer changes (i.e. converges). For example, after multiple cycles of adjustment, the candidate sentiment polarity prior information of the unlabeled customer service text sample E stabilizes at "positive", then this "positive" is the sentiment polarity prior information that the unlabeled customer service text sample E finally corresponds to. By executing step S40 of this implementation method, the cloud server can use the relationship network (text topology) between customer service text samples to effectively transmit the sentiment polarity prior information of the labeled customer service text samples to the unlabeled customer service text samples, thereby determining the sentiment polarity prior information of the unlabeled customer service text samples. This is of great significance to the entire artificial intelligence-based RPA customer service interaction optimization process. For example, accurate sentiment polarity prior information of unlabeled samples can improve the accuracy of sentiment polarity analysis, thereby enabling RPA customer service to provide appropriate response content based on more accurate sentiment information and improve customer satisfaction.

[0122] As an implementation method, the text topology structure includes text topology units corresponding to each customer service text sample and line segments between the text topology units, where the line segments represent the relationship between two connected text topology units. Step S403, combining the text topology structure, conducts prior information transmission on the sentiment polarity prior information of the labeled customer service text sample to obtain the sentiment polarity prior information corresponding to the unlabeled customer service text sample, which may specifically include:

[0123] Step S4031: using line segments between text topology units in the text topology structure to conduct prior information transmission on the sentiment polarity prior information of the labeled customer service text sample, so as to determine candidate sentiment polarity prior information of the unlabeled customer service text sample;

[0124] Step S4032: For each unlabeled customer service text sample, in combination with the text topology units adjacent to the unlabeled customer service text sample in the text topology structure, the candidate sentiment polarity prior information of the unlabeled customer service text sample is cyclically adjusted until the prior information of the text topology unit in the text topology structure no longer changes, thereby obtaining the sentiment polarity prior information corresponding to the unlabeled customer service text sample.

[0125] In the RPA customer service interaction scenario, the text topology structure constructed by the cloud server is established based on the relationship between customer service text samples. Each customer service text sample corresponds to a text topology unit, and the line segments between the units represent the relationship between them. This relationship can be quantified by the previous steps (such as the sample matching score calculated in step S401). For example, in the context of an e-commerce customer service, the customer service text samples may involve different types of content such as product evaluation, logistics consultation, and after-sales service.

[0126] Suppose customer service text sample A is about product evaluation, with the content "This product is very practical and cost-effective", sample B is about logistics consultation, with the content "My package has been shipped, when will it arrive?" Sample C is about after-sales service, with the content "The product had a small problem, and the after-sales staff solved it quickly." By calculating the relationship between sample description vectors (such as cosine similarity, etc.), it is found that the relationship between samples A and B is relatively distant, and the relationship between samples A and C is relatively close. In the text topology structure, this will be reflected in the line segment between A and C being "thick" (indicating the closeness of the relationship) greater than the line segment between A and B.

[0127] The labeled customer service text samples carry predetermined sentiment polarity prior information. For example, the sentiment polarity prior information of sample A is "positive" because it gives a positive evaluation of the product's features and cost-effectiveness. In the text topology, the cloud server uses the line segments connecting the labeled samples and the unlabeled samples to transmit this sentiment polarity prior information.

[0128] Suppose there is an unlabeled customer service text sample D, whose content is "The appearance design of this product is good." If in the text topology structure, the line segment between sample D and sample A indicates that they have a certain association (for example, the relationship determined by the sample matching score), then the "positive" sentiment polarity prior information of sample A will be transmitted to sample D, making the candidate sentiment polarity prior information of sample D tend to be "positive." The specific transmission strength can be determined according to the relationship strength represented by the line segment (such as the value of the sample matching score). If the sample matching score of sample D and sample A is high, then the impact of this transmission is greater, and sample D is more likely to be determined as the candidate sentiment polarity prior information of "positive."

[0129] For another example, the sentiment polarity prior information of labeled customer service text sample E is "negative", and the content is "the quality of this product is not very good." If the unlabeled customer service text sample F is connected to sample E in the text topology structure, even if the content of sample F does not explicitly express sentiment, such as "the color of the product is a bit special", due to its association with sample E, it will be affected by the "negative" sentiment polarity prior information of sample E, and its candidate sentiment polarity prior information will shift in the "negative" direction. The specific degree of shift depends on the strength of the relationship represented by the line segment between them.

[0130] For each unlabeled customer service text sample, the cloud server initially determines its candidate sentiment polarity prior information according to step S4031. Taking a hotel customer service scenario as an example, the content of the unlabeled customer service text sample G is "the hotel's room layout is very reasonable". In step S4031, due to its association with the labeled customer service text sample H (the sentiment polarity prior information is "positive", and the content is "the hotel's service is very thoughtful and the room is very comfortable"), its candidate sentiment polarity prior information is "positive". However, sample G also has relationships with other adjacent text topological units, and these relationships will be considered in subsequent cyclic adjustments.

[0131] The cloud server will adjust the candidate sentiment polarity prior information of the unlabeled customer service text sample based on the text topology units adjacent to it in the text topology structure. Assume that in addition to being adjacent to sample H, the unlabeled customer service text sample G is also adjacent to the unlabeled customer service text sample I (whose candidate sentiment polarity prior information is "neutral" and the content is "the hotel's breakfast variety is average"), and is adjacent to the labeled customer service text sample J (whose sentiment polarity prior information is "negative" and the content is "the hotel's price is a bit high").

[0132] In the first round of adjustments, the cloud server will comprehensively consider the relationship strength between G and these adjacent units (represented by line segments, based on previous sample matching scores, etc.) and their sentiment polarity prior information or candidate sentiment polarity prior information. If the relationship strength between G and H is 0.8 (assuming that this is a quantitative relationship strength based on the sample matching score), the relationship strength with I is 0.5, and the relationship strength with J is 0.3. Then during the adjustment, since the relationship between G and H is strong and H is "positive", this will have a greater driving effect on G's candidate sentiment polarity prior information in the "positive" direction; while the relationship between G and J is weak, but J is "negative", it will have a certain pulling effect on G's candidate sentiment polarity prior information in the "negative" direction; the relationship strength between G and I is moderate and I is "neutral", which will also have a certain impact on G's candidate sentiment polarity prior information. After combining these influences, G's candidate sentiment polarity prior information may be adjusted slightly from the initial "positive" to the "neutral" direction, for example, adjusted to a value between "positive" and "neutral".

[0133] The cloud server will continue to repeat this adjustment process. In each round of adjustment, the relationship between the unlabeled customer service text sample and all adjacent units will be reconsidered, because during the adjustment process, the sentiment polarity prior information or candidate sentiment polarity prior information of the adjacent units may also change. For example, in the second round of adjustment, the candidate sentiment polarity prior information of the unlabeled customer service text sample I may change from "neutral" to a value closer to "negative" due to the interaction with other adjacent units. This will in turn affect the adjustment results of the unlabeled customer service text sample G.

[0134] This cyclic adjustment process will continue until the prior information of all text topology units in the text topology structure no longer changes (i.e. converges). When the convergence state is reached, the final sentiment polarity prior information determined by the unlabeled customer service text sample G may be a relatively stable value, for example, "neutral to positive" after multiple rounds of adjustments. For the unlabeled customer service text sample, this finalized sentiment polarity prior information can more accurately reflect its sentiment tendency in the entire customer service text sample library, providing a more accurate basis for subsequent sentiment polarity analysis, RPA customer service response optimization and other operations.

[0135] In this way, the cloud server effectively uses the text topology structure to conduct and adjust the sentiment polarity prior information of the labeled customer service text samples in step S403, thereby determining the sentiment polarity prior information of the unlabeled customer service text samples. This helps to improve the accuracy and effectiveness of the entire artificial intelligence-based RPA customer service interaction optimization process, allowing RPA customer service to provide customers with more appropriate response content based on more accurate sentiment information, thereby improving customer satisfaction.

[0136] As an implementation mode, step S50, for each text paragraph of the customer service text sample, combines the paragraph description vector corresponding to the text paragraph and the sample description vector of the customer service text sample corresponding to the text paragraph to obtain a semantic unity measure between the text paragraph and its corresponding customer service text sample, which may specifically include:

[0137] Step S51: establishing a sample description vector library based on the sample description vectors of each customer service text sample, and establishing a paragraph description vector library based on the paragraph description vectors of the text paragraphs of each customer service text sample;

[0138] Step S52: for each customer service text sample, index at least one target sample description vector similar to the sample description vector of the customer service text sample in the sample description vector library, and determine a set of pending sentiment polarity prior information of the customer service text sample based on the sentiment polarity prior information of the customer service text sample corresponding to the target sample description vector;

[0139] Step S53: for each text paragraph of the customer service text sample, index at least one target paragraph description vector similar to the paragraph description vector of the text paragraph in the paragraph description vector library, and determine a set of pending sentiment polarity prior information of the text paragraph based on the sentiment polarity prior information of the customer service text sample corresponding to the target paragraph description vector;

[0140] Step S54: Obtain a semantic unity measure between the text paragraph and its corresponding customer service text sample based on the pending sentiment polarity prior information set of the text paragraph and the pending sentiment polarity prior information set of the customer service text sample corresponding to the text paragraph.

[0141] In the RPA customer service interaction, the cloud server has obtained the sample description vectors of each customer service text sample through the previous steps (such as step S30). These sample description vectors are a quantitative representation of the entire customer service text sample, which contains various feature information related to sentiment polarity analysis. For example, in an e-commerce customer service scenario, for the customer service text sample "This product is of good quality, but the price is a bit high, but the customer service attitude is very good", its sample description vector may be [0.7, 0.3, 0.8] (here it is assumed that the three dimensions represent the quantitative evaluation of product quality, price, and customer service attitude respectively).

[0142] The cloud server collects sample description vectors of all customer service text samples, and then builds a sample description vector library. This library is like a warehouse that stores the feature information of all customer service text samples. For example, if there are multiple customer service text samples, such as "The product's functions are very practical, and the logistics speed is also fast" (the sample description vector is [0.8, 0.7, 0], assuming that the third dimension represents other aspects not involved), "The product's appearance is not very good, but the cost performance is high" (the sample description vector is [0.3,0.8, 0]), etc., these sample description vectors will be stored in the sample description vector library.

[0143] For each customer service text sample, the cloud server will also perform text splitting (such as step S20) to obtain multiple text paragraphs, and there will be a corresponding paragraph description vector for each text paragraph (such as step S30). For example, for the above-mentioned customer service text sample "The quality of this product is good, but the price is a bit high, but the customer service attitude is very good", the split text paragraphs may be "The quality of this product is good", "But the price is a bit high", "But the customer service attitude is very good". The paragraph description vector of the text paragraph "The quality of this product is good" may be [0.8, 0, 0] (focusing on the product quality dimension). The cloud server also collects the paragraph description vectors of all these text paragraphs and establishes a paragraph description vector library. This library will store the feature information of all text paragraphs and provide a data basis for subsequent analysis.

[0144] For each customer service text sample, the cloud server searches for a target sample description vector similar to its sample description vector in the sample description vector library. The method for determining similarity can be to calculate the cosine similarity between the vectors (such as the method that may be used in other steps before).

[0145] Once the customer service text sample corresponding to the target sample description vector is determined, the cloud server determines the pending emotional polarity prior information set of the current customer service text sample based on the emotional polarity prior information of these customer service text samples. For example, the emotional polarity prior information of the customer service text sample "The product has great functions and is easy to operate" corresponding to the target sample description vector found above is "positive". If other similar customer service text samples are found, and their emotional polarity prior information is also "positive", then the pending emotional polarity prior information set of the current customer service text sample "This product has powerful functions and is easy to use" is {"positive"}. If a similar customer service text sample with emotional polarity prior information of "mixed (both positive and negative)" is found, then the pending emotional polarity prior information set becomes {"positive", "mixed"}.

[0146] For each text paragraph of the customer service text sample, the cloud server performs a similar index operation in the paragraph description vector library. For example, for the text paragraph "This product is of good quality, but the price is a bit high, but the customer service attitude is very good" in the customer service text sample, its paragraph description vector is [0.8, 0, 0]. The cloud server calculates the similarity between this vector and other paragraph description vectors in the paragraph description vector library.

[0147] Assume that there is a text paragraph in the library, "The quality of the product is quite good", and its paragraph description vector is [0.9, 0, 0]. Calculate their cosine similarity (calculation method is the same as above). If the similarity threshold requirement is met, the text paragraph corresponding to the vector is determined as the text paragraph corresponding to the target paragraph description vector.

[0148] Then, the cloud server determines the pending sentiment polarity prior information set of the text paragraph based on the sentiment polarity prior information of the customer service text sample corresponding to the target paragraph description vector. For example, if the sentiment polarity prior information of the customer service text sample "The quality of the product is quite good" corresponding to the target paragraph description vector is "positive", then the pending sentiment polarity prior information set of the text paragraph "The quality of this product is good" is {"positive"}. If multiple similar customer service text samples corresponding to the target paragraph description vector are found, and their sentiment polarity prior information includes "positive" and "mixed", then the pending sentiment polarity prior information set of the text paragraph becomes {"positive", "mixed"}.

[0149] The cloud server first determines the number of occurrences of the sentiment polarity prior information in the pending sentiment polarity prior information set of the text paragraph, thereby determining the sentiment polarity probability distribution of the text paragraph. For example, for the text paragraph "The quality of this product is good", its pending sentiment polarity prior information set is {"positive", "mixed"}. Assuming that "positive" appears once and "mixed" appears once in the set, the sentiment polarity probability distribution of the text paragraph is positive probability. , the mixing probability , negative probability (Since there is no "negative" in the set).

[0150] Similarly, the cloud server determines the number of occurrences of the sentiment polarity prior information in the pending sentiment polarity prior information set of the customer service text sample corresponding to the text paragraph, and determines the sentiment polarity probability distribution of the customer service text sample. For example, the pending sentiment polarity prior information set of the customer service text sample "This product is of good quality, but the price is a bit high, but the customer service attitude is very good" is {"positive", "mixed"}. Assuming that "positive" appears 3 times and "mixed" appears 2 times, then the sentiment polarity probability distribution of the customer service text sample is positive probability. , the mixing probability .4, Negative probability .

[0151] A method for calculating a semantic unity measure may be to calculate the Euclidean distance of the probability distribution of the two, and the distance value may be used as a measure of the semantic unity between a text paragraph and its corresponding customer service text sample, and the smaller the distance, the higher the semantic unity. For example, if the calculation result of the semantic unity measure of another text paragraph and its corresponding customer service text sample is 0.1, then the previously calculated 0.02 indicates that the former has a higher semantic unity. Through the execution of step S50 of this embodiment, the cloud server can deeply analyze the relationship between the customer service text sample and its text paragraph from a semantic perspective, which is of great significance for subsequent sentiment polarity analysis, RPA customer service response optimization and other links. For example, a semantic unity measure can help the cloud server more accurately judge the sentiment polarity of the entire customer service text sample, avoid one-sided judgment based only on a certain text paragraph, so that RPA customer service can provide appropriate response content based on more accurate sentiment information, and improve customer satisfaction.

[0152] As an implementation mode, step S54, based on the pending sentiment polarity prior information set of the text paragraph and the pending sentiment polarity prior information set of the customer service text sample corresponding to the text paragraph, obtains the semantic unity measure between the text paragraph and its corresponding customer service text sample, which may specifically include:

[0153] Step S541: determining the number of occurrences of the sentiment polarity prior information in the pending sentiment polarity prior information set of the text paragraph, and determining the sentiment polarity probability distribution of the text paragraph;

[0154] Step S542: determining the number of occurrences of the sentiment polarity prior information in the pending sentiment polarity prior information set of the customer service text sample corresponding to the text paragraph, and determining the sentiment polarity probability distribution of the customer service text sample corresponding to the text paragraph;

[0155] Step S543: Obtain a semantic unity measure between the text paragraph and its corresponding customer service text sample based on the sentiment polarity probability distribution of the text paragraph and the sentiment polarity probability distribution of the customer service text sample corresponding to the text paragraph.

[0156] In step S541, the task of the cloud server is to determine the number of occurrences of the sentiment polarity prior information in the pending sentiment polarity prior information set of the text paragraph. Assume that the cloud server already has a pending sentiment polarity prior information set of the text paragraph, for example, this set may contain three types of sentiment polarity prior information, namely "positive", "negative" and "neutral". The cloud server needs to count the number of times each type of sentiment polarity prior information appears in this set.

[0157] This statistic is similar to building a simple frequency histogram to represent the distribution of data. The sentiment polarity prior information can be regarded as different categories, and the number of occurrences is the frequency of each category. For example, in the application scenario of a simple naive Bayes classifier, if the sentiment polarity classification is regarded as a classification task, this step of determining the probability distribution of sentiment polarity is similar to calculating the prior probability. Suppose there is a naive Bayes model for sentiment polarity classification, and the sentiment polarity is divided into three categories: "positive" P, "negative" N, and "neutral" Z. If there are a total of n = 5 elements in the set of pending sentiment polarity prior information for a certain text paragraph, of which "positive" appears 3 times, "negative" appears 1 time, and "neutral" appears 1 time. Then the probability of "positive" is , the probability of "negative" , the probability of "neutral" . This is the probability distribution of sentiment polarity for this text paragraph.

[0158] In step S542, the cloud server determines the number of occurrences of the sentiment polarity prior information in the pending sentiment polarity prior information set of the customer service text sample corresponding to the text paragraph. Similar to step S541, the operation object becomes the pending sentiment polarity prior information set of the customer service text sample corresponding to the text paragraph.

[0159] In step S543, the cloud server obtains a semantic unity metric based on the sentiment polarity probability distribution of the text paragraph obtained above and the sentiment polarity probability distribution of the customer service text sample corresponding to the text paragraph. This metric can reflect the consistency of the text paragraph and its corresponding customer service text sample in terms of sentiment polarity.

[0160] One way to calculate a semantic unity measure is to calculate the distance between two probability distributions. For example, the KL divergence can be used to measure the difference between two probability distributions P and P'. The calculation formula for KL divergence is , where i represents the category of sentiment polarity (such as positive, negative, neutral, etc.). However, it should be noted that the KL divergence is asymmetric, that is, Other distance metrics can also be used, such as Euclidean distance. If the sentiment polarity probability distribution is regarded as a vector, for example, the sentiment polarity probability distribution of a text paragraph is (Assuming there are only three sentiment polarity categories, is the probability of positive sentiment polarity, is the probability of negative sentiment polarity, is the probability of neutral sentiment polarity), for the sentiment polarity probability distribution of customer service text samples , Euclidean distance The smaller the distance value is, the higher the semantic unity between the text paragraph and its corresponding customer service text sample is.

[0161] As an implementation mode, step S60, based on the semantic unity metric and the sentiment polarity prior information corresponding to the unlabeled customer service text sample, updates the network weights of the basic sentiment polarity analysis network to obtain a converged sentiment polarity analysis network, which may specifically include:

[0162] Step S61: performing sentiment polarity analysis on the customer service text sample according to the sample description vector of the customer service text sample to obtain a sample sentiment polarity probability distribution corresponding to the customer service text sample;

[0163] Step S62: performing sentiment polarity analysis on the text paragraph according to the paragraph description vector corresponding to the text paragraph, and obtaining a paragraph sentiment polarity probability distribution corresponding to the text paragraph;

[0164] Step S63: combining the sample sentiment polarity probability distribution, the paragraph sentiment polarity probability distribution, the semantic unity measure and the sentiment polarity prior information corresponding to the unlabeled customer service text sample, the network weights of the basic sentiment polarity analysis network are updated to obtain a converged sentiment polarity analysis network.

[0165] The cloud server performs sentiment polarity analysis on the customer service text sample based on the sample description vector of the customer service text sample. The sample description vector here is mined in the previous step (such as step S30), and it contains the representation information of the customer service text sample in a certain feature space. The cloud server uses the basic sentiment polarity analysis network (which can be a neural network model, such as a multi-layer perceptron model) to process this sample description vector. In this process, the output of the network is the probability distribution of various sentiment polarities (such as positive, negative, neutral, etc.) corresponding to the customer service text sample. This probability distribution represents the probability of the customer service text sample belonging to each sentiment polarity.

[0166] Assume that the output layer of the network has three neurons, corresponding to positive, negative, and neutral sentiment polarities. The output result may be ([0.6, 0.2, 0.2]), which means that the probability of this customer service text sample being judged as positive sentiment polarity is 0.6, the probability of negative sentiment polarity is 0.2, and the probability of neutral sentiment polarity is 0.2. This probability distribution is the sample sentiment polarity probability distribution corresponding to the customer service text sample.

[0167] Taking the Naive Bayes classifier as an example, if the sentiment polarity classification is regarded as a classification task, then the features in the sample description vector are similar to the attributes in Naive Bayes. The Naive Bayes classifier calculates the posterior probability that the sample belongs to each category based on the features in the sample description vector, the prior probability and conditional probability obtained in the training phase, which is similar to the sample sentiment polarity probability distribution obtained here. For neural network models, such as the multi-layer perceptron mentioned above, it calculates the output sentiment polarity probability distribution from the input sample description vector through a series of linear transformations and nonlinear activation functions (such as the ReLU function).

[0168] In step S62, sentiment polarity analysis is performed on the text paragraph based on the paragraph description vector corresponding to the text paragraph. Similar to step S61, the operation object is changed to the paragraph description vector of the text paragraph. Each text paragraph has its corresponding paragraph description vector, which also contains feature information such as semantics and structure of the paragraph. The cloud server inputs the paragraph description vector into the basic sentiment polarity analysis network, and the network outputs the probability distribution of various sentiment polarities corresponding to the text paragraph.

[0169] Assume that the probability distribution of the network output is [0.8, 0.1, 0.1], which means that the probability of this text paragraph being judged as positive sentiment polarity is 0.8, the probability of negative sentiment polarity is 0.1, and the probability of neutral sentiment polarity is 0.1. This probability distribution is the probability distribution of paragraph sentiment polarity corresponding to the text paragraph.

[0170] In step S63, the cloud server updates the network weights of the basic sentiment polarity analysis network by combining the sample sentiment polarity probability distribution, the paragraph sentiment polarity probability distribution, the semantic unity metric, and the sentiment polarity prior information corresponding to the unlabeled customer service text samples.

[0171] First, the semantic unity metric reflects the degree of consistency between text paragraphs and customer service text samples in terms of sentiment polarity. The sentiment polarity prior information corresponding to the unlabeled customer service text samples provides additional information for the model. The cloud server uses this information and the sentiment polarity probability distribution of samples and paragraphs to adjust the network weights, so that the network can better adapt to the data and improve the accuracy of sentiment polarity analysis.

[0172] When updating the network weights, some optimization algorithms can be used, such as stochastic gradient descent (SGD) or its variants (such as Adagrad, Adadelta, etc.). These algorithms adjust the network weights according to a certain cost function (loss function) to minimize the cost function.

[0173] For example, suppose that in an RPA customer service interaction scenario of an online travel platform, there are a batch of customer service text samples and their split text paragraphs. For a customer service text sample, its sample sentiment polarity probability distribution is Ps = [0.7, 0.2, 0.1] (corresponding to positive, negative, and neutral, respectively), the paragraph sentiment polarity probability distribution of one text paragraph is (Pp = [0.9, 0.05, 0.05]), the semantic unity metric (assuming that it is calculated using the Euclidean distance) is (d = 0.2), and the sentiment polarity prior information corresponding to the unlabeled customer service text sample (assuming that it is obtained through the previous feature clustering operation) indicates that this sample is more likely to have positive sentiment polarity.

[0174] Assume that the basic sentiment polarity analysis network is a multi-layer neural network, and the network weights are represented by W. The cloud server uses the stochastic gradient descent algorithm to update the weights. First, a cost function J needs to be defined. This cost function can comprehensively consider the probability distribution of sample sentiment polarity, the probability distribution of paragraph sentiment polarity, the semantic unity metric, and the sentiment polarity prior information corresponding to the unlabeled customer service text sample. For example, ,in is the true sentiment polarity label of the sample (if any), is the true sentiment polarity label of the paragraph (if any), is the sentiment polarity prior information corresponding to the unlabeled customer service text sample, and is the weight coefficient, Is a custom function for comprehensive consideration The relationship between.

[0175] Then, the network weight W is continuously updated iteratively according to the stochastic gradient descent algorithm until the cost function J converges. When the cost function converges, the network at this time is a converged sentiment polarity analysis network.

[0176] As an implementation method, step S63, combining the sample sentiment polarity probability distribution, the paragraph sentiment polarity probability distribution, the semantic unity measure and the sentiment polarity prior information corresponding to the unlabeled customer service text sample, updates the network weights of the basic sentiment polarity analysis network to obtain a converged sentiment polarity analysis network, which may specifically include:

[0177] Step S631: for each customer service text sample, based on the paragraph sentiment polarity probability distribution of each text paragraph in the customer service text sample and the semantic unity measurement corresponding to each text paragraph, the sentiment polarity prior information of the customer service text sample is optimized to obtain the optimized sentiment polarity prior information of the customer service text sample;

[0178] Step S632: Integrate the optimized sentiment polarity prior information of each customer service text sample according to the sample sentiment polarity probability distribution corresponding to each customer service text sample to obtain a fusion training cost;

[0179] Step S633: based on the fusion training cost, the network weights of the basic sentiment polarity analysis network are updated to obtain a converged sentiment polarity analysis network.

[0180] In step S631, the cloud server optimizes the sentiment polarity prior information of each customer service text sample by using the paragraph sentiment polarity probability distribution of each text paragraph and the semantic unity metric corresponding to each text paragraph. For each customer service text sample, it contains multiple text paragraphs, each of which has its own paragraph sentiment polarity probability distribution, which represents the possibility that the paragraph belongs to different sentiment polarities (such as positive, negative, and neutral). At the same time, the semantic unity metric reflects the degree of consistency between the text paragraph and its corresponding customer service text sample in terms of sentiment polarity.

[0181] The cloud server adjusts the sentiment polarity prior information of the customer service text sample by comprehensively considering this information, so that it more accurately reflects the sentiment polarity of the customer service text sample. This optimization is based on the assumption that if the paragraph sentiment polarity probability distribution of a text paragraph is combined with the semantic unity measure between the paragraph and the customer service text sample, it can provide a more accurate estimate of the sentiment polarity prior information of the entire customer service text sample.

[0182] For example, consider an RPA customer service interaction scenario of an online shopping platform. There is a customer service text sample about customer feedback on purchased electronic products, and this customer service text sample is divided into three text paragraphs.

[0183] For the first text paragraph, its paragraph sentiment polarity probability distribution is: (representing the probabilities of positive, negative, and neutral sentiment polarities, respectively), and the semantic unity measure between it and the customer service text sample (assuming it is obtained by the previous calculation method) is For the second text paragraph, , ; For the third text paragraph, .

[0184] Assume that the initial sentiment polarity prior information of this customer service text sample is that it is more likely to be positive sentiment polarity (for example, the probability is 0.7). The cloud server will optimize this prior information based on the information of each paragraph. For example, the cloud server may combine the paragraph sentiment polarity probability distribution and the semantic unity measure based on a weighted average method to calculate the adjustment amount to the prior information. If the adjustment weight of the prior information for the first paragraph is w1, the second paragraph is w2, and the third paragraph is w3 (these weights can be determined based on factors such as the importance of the paragraph, for example, based on the length of the paragraph or its relevance to the topic), then the optimized sentiment polarity prior information This can be calculated as follows (this is just a simple example): ,in is the initial sentiment polarity prior information, and f is a custom function used to distribute the sentiment polarity probability of the paragraph. and semantic unity measure Calculate the adjustment. Assume w1=0.3, w2=0.3, w3=0.4, and the calculation method of f function is (here is the probability of positive sentiment polarity in the paragraph sentiment polarity probability distribution), then for the first paragraph , for the second paragraph , for the third paragraph .but , this is the sentiment polarity prior information of the optimized customer service text sample.

[0185] In step S632, the cloud server integrates the optimized sentiment polarity prior information of each customer service text sample according to the sample sentiment polarity probability distribution corresponding to each customer service text sample, thereby obtaining a fusion training cost. For each customer service text sample, the cloud server first performs a specific operation on its sample sentiment polarity probability distribution (the example given here is a logarithmic operation), and then integrates the logarithmic operation result and the optimized sentiment polarity prior information of the customer service text sample to obtain the sample sub-cost of the customer service text sample. Finally, the sample sub-costs of each customer service text sample are integrated to obtain the fusion training cost.

[0186] This fusion training cost is a comprehensive indicator that reflects the overall situation of the entire customer service text sample set after considering the sample sentiment polarity probability distribution and the optimized sentiment polarity prior information. This cost will be used in the subsequent network weight update process.

[0187] Let’s continue to use the RPA customer service interaction scenario of the online shopping platform as an example. Assume that there are three customer service text samples, namely .for , the probability distribution of sample sentiment polarity is , optimized sentiment polarity prior information First of all, Perform logarithmic operations, assuming natural logarithm ,get . Then compare it with Integration (assuming that the integration method is weighted summation, the weights are , get the sample cost .

[0188] Likewise, for , assuming , , calculate the sample cost .for , assuming , , calculate the sample cost .

[0189] Finally, Perform integration processing (assuming simple summation) to obtain the fusion training cost .

[0190] The training cost is an indicator that measures the difference between the model's prediction results and the actual results. The fusion training cost constructed here is similar to a custom loss function. In neural network training, common loss functions include cross entropy loss function and mean square error loss function. The fusion training cost constructed here comprehensively considers the probability distribution of sample sentiment polarity and the optimized sentiment polarity prior information. This custom loss function can better adapt to the sentiment polarity analysis task in the RPA customer service interaction scenario. By minimizing this fusion training cost, the network weights can be adjusted to make the model's prediction results more consistent with the actual situation. For example, when using optimization algorithms such as stochastic gradient descent to train a neural network, the network weights are updated with the goal of minimizing the loss function.

[0191] In step S633, the cloud server updates the network weights of the basic sentiment polarity analysis network based on the obtained fusion training cost. In this process, the fusion training cost is used as an indicator to measure the performance of the model, and the cloud server uses an optimization algorithm (such as stochastic gradient descent or its variants) to adjust the network weights. The optimization algorithm determines the update direction and step size of the weights based on the gradient of the fusion training cost to the network weights. As the weights are continuously updated, the fusion training cost will gradually decrease. When the fusion training cost converges (reaches a preset threshold or no longer changes significantly), the network at this time is a converged sentiment polarity analysis network.

[0192] This converged network can perform sentiment polarity analysis on RPA customer service conversation text more accurately because its weights are optimized by comprehensively considering the sample sentiment polarity probability distribution, paragraph sentiment polarity probability distribution, semantic unity measurement, and sentiment polarity prior information corresponding to unlabeled customer service text samples.

[0193] As an implementation mode, step S633, based on the fusion training cost, updates the network weights of the basic sentiment polarity analysis network to obtain a converged sentiment polarity analysis network, which may specifically include:

[0194] Step S6331: Obtaining the paragraph classification cost of the text paragraph according to the paragraph sentiment polarity probability distribution of the text paragraph and the sentiment polarity prior information of the customer service text sample corresponding to the text paragraph;

[0195] Step S6332: Based on the fusion training cost and the paragraph classification cost, the network weights of the basic sentiment polarity analysis network are updated to obtain a converged sentiment polarity analysis network.

[0196] In step S6331, the cloud server obtains the paragraph classification cost based on the paragraph sentiment polarity probability distribution of the text paragraph and the sentiment polarity prior information of the customer service text sample corresponding to the text paragraph. The paragraph sentiment polarity probability distribution indicates the possibility that the text paragraph belongs to different sentiment polarities (such as positive, negative, and neutral), while the sentiment polarity prior information of the customer service text sample is the prior knowledge about the sentiment polarity of the entire customer service text sample that has been optimized or determined before. The paragraph classification cost is an indicator that measures the difference between the sentiment polarity classification of a text paragraph and the prior information. The calculation of this cost can be based on different methods, such as a distance metric, which is used to reflect the gap between the paragraph sentiment polarity probability distribution and the probability distribution expected based on the prior information.

[0197] For example, a customer service text sample is about customer feedback on a new package. This customer service text sample is split into several text paragraphs. For one of the text paragraphs, the probability distribution of its paragraph sentiment polarity is (corresponding to the probabilities of positive, negative, and neutral sentiment polarities, respectively), and the sentiment polarity prior information of the customer service text sample corresponding to this text paragraph is that there is a higher probability of positive sentiment polarity (assuming the probability is 0.8).

[0198] One method to calculate the paragraph classification cost can be the mean square error (MSE). Assume that the expected probability distribution based on prior information is , then the paragraph classification cost =0.02.

[0199] In step S6332, the cloud server updates the network weights of the basic sentiment polarity analysis network based on the fusion training cost and the paragraph classification cost. The fusion training cost is an overall training cost obtained by comprehensively considering the sample sentiment polarity probability distribution of each customer service text sample, the optimized sentiment polarity prior information and other factors. It reflects the performance of the entire model when processing the customer service text sample set. The paragraph classification cost is a measure of the classification difference between each text paragraph and the prior information.

[0200] By combining these two costs, the performance of the model at different levels (overall samples and individual text paragraphs) can be considered more comprehensively, so that the network weights can be updated more accurately. The cloud server can use optimization algorithms such as stochastic gradient descent (SGD) or its variants to adjust the network weights according to the combined impact of these two costs. As the weights are updated, the performance of the model will gradually improve, and the fusion training cost and paragraph classification cost will gradually decrease. When they reach convergence (for example, the change in cost is less than a preset threshold), the network at this time is a converged sentiment polarity analysis network.

[0201] As an implementation mode, step S6332, based on the fusion training cost and the paragraph classification cost, updates the network weights of the basic sentiment polarity analysis network to obtain a converged sentiment polarity analysis network, which may specifically include:

[0202] Step S63321: performing deviation analysis between the paragraph sentiment polarity probability distribution of the text paragraph and the equal probability distribution to obtain a deviation value corresponding to the text paragraph;

[0203] Step S63322: According to the semantic unity metric corresponding to the text paragraph, the paragraph classification cost and the deviation value are integrated to obtain the paragraph cost of the text paragraph;

[0204] Step S63323: combining the fusion training cost and the paragraph cost corresponding to each text paragraph of the customer service text sample, updating the network weights of the basic sentiment polarity analysis network to obtain a converged sentiment polarity analysis network.

[0205] In step S63321, the cloud server performs a deviation analysis between the paragraph sentiment polarity probability distribution and the equal probability distribution of the text paragraph. The paragraph sentiment polarity probability distribution indicates the probability of a text paragraph belonging to different sentiment polarities (such as positive, negative, and neutral). The equal probability distribution is a distribution that assumes that all sentiment polarities have the same probability. By analyzing the deviation between the two, a deviation value reflecting the deviation of the sentiment polarity probability distribution of the text paragraph from uniformity can be obtained. This deviation value can be calculated using a variety of methods, for example, the degree of difference between the paragraph sentiment polarity probability distribution and the equal probability distribution can be measured based on a certain distance metric or difference metric. This deviation value reflects the degree of difference between the sentiment polarity probability distribution of the text paragraph and the equal probability distribution. In machine learning, this deviation value is similar to a characteristic measure of a data distribution. For example, in a clustering algorithm, when analyzing the degree of deviation of a data point from a certain center (the equal probability distribution here can be regarded as a special center), a similar distance metric is used to determine the attribution of the data point or evaluate the effect of clustering. For the sentiment polarity analysis network, this deviation value can provide additional information about the distribution characteristics of the sentiment polarity of the text paragraph, which helps to update the network weights more accurately.

[0206] In step S63322, the cloud server integrates the paragraph classification cost and the deviation value according to the semantic unity metric corresponding to the text paragraph, thereby obtaining the paragraph cost of the text paragraph. The semantic unity metric reflects the degree of consistency between the text paragraph and its corresponding customer service text sample in terms of sentiment polarity. The paragraph classification cost is an indicator that measures the difference between the sentiment polarity classification of the text paragraph and the prior information, and the deviation value is the deviation between the probability distribution of the sentiment polarity of the paragraph and the equal probability distribution. By integrating these three factors, a paragraph cost that comprehensively considers multiple features of the text paragraph can be obtained. The integration method can be a weighted summation or other custom function form, the purpose of which is to integrate these three different aspects of information into a single metric for subsequent use in updating network weights. For example, continuing with the above-mentioned RPA customer service interaction scenario in the financial services field. For that text paragraph, assuming the paragraph classification cost (previously calculated based on the probability distribution of paragraph sentiment polarity and sentiment polarity prior information of customer service text samples), semantic unity metric (Assumed to be obtained through previous calculation).

[0207] Assume that the integration method is weighted summation, and the weights are Used for paragraph classification cost, semantic unity measure and deviation value. Then the paragraph cost Calculation can be obtained . This paragraph cost comprehensively considers the text paragraph's sentiment polarity classification accuracy, semantic consistency with customer service text samples, and the characteristics of its own sentiment polarity probability distribution. In machine learning, this paragraph cost is similar to a custom loss function that combines multiple factors related to model performance. For example, in the training of deep learning models, multiple loss terms (such as classification loss, regularization loss, etc.) are sometimes combined into a total loss function. The paragraph cost here is similar to such a combined loss function, but it targets multiple features of a single text paragraph.

[0208] In step S63323, the cloud server updates the network weights of the basic sentiment polarity analysis network by combining the fusion training cost and the paragraph cost corresponding to each text paragraph of the customer service text sample. The fusion training cost is an overall training cost obtained by comprehensively considering factors such as the sample sentiment polarity probability distribution of each customer service text sample and the optimized sentiment polarity prior information, while the paragraph cost is a cost obtained by combining multiple features for each text paragraph.

[0209] By combining the two, the performance of the model at the level of overall samples and individual text paragraphs can be comprehensively considered. The cloud server uses an optimization algorithm (such as stochastic gradient descent or its variants) to adjust the network weights according to this comprehensive cost. As the weights are continuously updated, the fusion training cost and paragraph cost will gradually decrease. When they reach convergence (for example, the change in cost is less than a preset threshold), the network at this time is a converged sentiment polarity analysis network.

[0210] For example, suppose the basic sentiment polarity analysis network is a multi-layer perceptron (MLP) network, and the network weights are represented by W. The fusion training cost is , for the n text paragraphs in the customer service text sample, their paragraph costs are .

[0211] The cloud server uses an improved stochastic gradient descent algorithm to update the weights. Assume that the update formula is ,in is the learning rate, It is the weight coefficient for each paragraph cost, which is used to balance the impact of fusion training cost and each paragraph cost in weight update.

[0212] For example, in an RPA customer service interaction scenario of an e-commerce platform, there are a batch of customer service text samples and their split text paragraphs. Assume that the fusion training cost , there are three text paragraphs, and the paragraph costs are , the learning rate is 0.01, and the weight coefficient is .

[0213] The cloud server first calculates and (This needs to be calculated according to the structure of the network and the specific form of the cost function. For example, for an MLP with m input neurons, h hidden neurons and k output neurons, the calculation process involves the back-propagation algorithm.

[0214] Then the network weights \(W\) are updated according to the update formula. After multiple iterations, as the weights are updated, the fusion training cost and the paragraph cost will gradually decrease. For example, after 100 iterations, the fusion training cost =1, the paragraph costs are When the iteration continues, the two costs no longer change significantly (for example, in 10 consecutive iterations, the cost change is less than 0.001), and the network at this time is a converged sentiment polarity analysis network. This converged network can more accurately perform sentiment polarity analysis on RPA customer service conversation texts because its network weights are optimized by comprehensively considering the fusion training cost and paragraph cost.

[0215] Figure 2 A schematic diagram of a hardware entity of a cloud server provided in an embodiment of the present application, such as Figure 2 As shown, the hardware entity of the cloud server 1000 includes: a processor 1001 and a memory 1002, wherein the memory 1002 stores a computer program that can be run on the processor 1001, and the processor 1001 implements the steps in the method of any of the above embodiments when executing the program.

[0216] The above is only an implementation method of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application.

Claims

1. An artificial intelligence-based RPA customer service interaction optimization method, characterized in that: include: Acquire a basic sentiment polarity analysis network and a customer service text sample library, wherein the customer service text samples in the customer service text sample library include labeled customer service text samples and unlabeled customer service text samples, and the labeled customer service text samples include corresponding sentiment polarity prior information; For each customer service text sample, the customer service text sample is split into text segments to obtain multiple text paragraphs of the customer service text sample; Using the basic sentiment polarity analysis network, respectively perform text description vector mining on the customer service text sample and each text paragraph of the customer service text sample to obtain a sample description vector of the customer service text sample and a paragraph description vector corresponding to each text paragraph; According to the sentiment polarity prior information of the labeled customer service text sample, a feature clustering operation is performed on the sample description vector of each customer service text sample in the customer service text sample library to obtain the sentiment polarity prior information corresponding to the unlabeled customer service text sample; For each text paragraph of the customer service text sample, combining the paragraph description vector corresponding to the text paragraph and the sample description vector of the customer service text sample corresponding to the text paragraph, obtaining a semantic unity measure between the text paragraph and its corresponding customer service text sample; According to the semantic unity metric and the sentiment polarity prior information corresponding to the unlabeled customer service text sample, the network weights of the basic sentiment polarity analysis network are updated to obtain a converged sentiment polarity analysis network; Get the RPA customer service conversation text; The convergent sentiment polarity analysis network is called based on the RPA customer service conversation text, and sentiment polarity analysis is performed on the RPA customer service conversation text based on the convergent sentiment polarity analysis network to obtain a sentiment polarity label of the RPA customer service conversation text; According to the sentiment polarity label of the RPA customer service conversation text, based on the preset customer service reply content mapping strategy, the RPA reply content is optimized; Among them, the method of performing feature clustering operation on the sample description vector of each customer service text sample in the customer service text sample library based on the sentiment polarity prior information of the labeled customer service text sample to obtain the sentiment polarity prior information corresponding to the unlabeled customer service text sample includes: Deploy a number of feature cluster centers, and obtain a matching score between a sample description vector of each customer service text sample in the customer service text sample library and each feature cluster center; For each customer service text sample, according to the matching score and the sentiment polarity prior information of the labeled customer service text sample, a target feature cluster center matching the customer service text sample is determined from among the feature cluster centers, and the customer service text sample is added to the customer service text library corresponding to the target feature cluster center; For the customer service text library corresponding to each feature cluster center, a customer service text sample that meets the feature cluster center condition is screened out from the customer service text library and is determined as a new feature cluster center; Repeat the process of obtaining the matching score between the sample description vector of each customer service text sample in the customer service text sample library and each feature cluster center, stop when the obtained feature cluster center meets the clustering cutoff condition, and determine the sentiment polarity prior information of the unlabeled customer service text samples in the target customer service text library based on the sentiment polarity prior information of the labeled customer service text samples in the target customer service text library corresponding to the feature cluster center that meets the clustering cutoff condition; Or, the method of performing feature clustering operation on the sample description vector of each customer service text sample in the customer service text sample library based on the sentiment polarity prior information of the labeled customer service text sample to obtain the sentiment polarity prior information corresponding to the unlabeled customer service text sample includes: Determine a sample matching score between each two customer service text samples based on the sample description vectors between each two customer service text samples; In combination with the sample matching scores, a text topology structure is established, wherein the text topology structure represents the relationship between each customer service text sample; Combined with the text topology, the sentiment polarity prior information of the labeled customer service text sample is transmitted to obtain the sentiment polarity prior information corresponding to the unlabeled customer service text sample; For each text paragraph of the customer service text sample, combining the paragraph description vector corresponding to the text paragraph and the sample description vector of the customer service text sample corresponding to the text paragraph, obtaining a semantic unity measure between the text paragraph and the corresponding customer service text sample, including: A sample description vector library is established based on the sample description vectors of each customer service text sample, and a paragraph description vector library is established based on the paragraph description vectors of the text paragraphs of each customer service text sample; For each customer service text sample, at least one target sample description vector similar to the sample description vector of the customer service text sample is indexed in the sample description vector library, and a pending emotional polarity prior information set of the customer service text sample is determined based on the emotional polarity prior information of the customer service text sample corresponding to the target sample description vector; For each text paragraph of the customer service text sample, index at least one target paragraph description vector similar to the paragraph description vector of the text paragraph in the paragraph description vector library, and determine a pending sentiment polarity prior information set of the text paragraph based on sentiment polarity prior information of the customer service text sample corresponding to the target paragraph description vector; According to the pending sentiment polarity prior information set of the text paragraph and the pending sentiment polarity prior information set of the customer service text sample corresponding to the text paragraph, a semantic unity measure between the text paragraph and its corresponding customer service text sample is obtained.

2. The method according to claim 1, characterized in that The text topology structure includes text topology units corresponding to each customer service text sample and line segments between the text topology units, wherein the line segments represent the relationship between two connected text topology units; the prior information transmission of the sentiment polarity prior information of the labeled customer service text sample is performed in combination with the text topology structure to obtain the sentiment polarity prior information corresponding to the unlabeled customer service text sample, including: In the text topology structure, the line segments between the text topology units are used to conduct prior information on the sentiment polarity prior information of the labeled customer service text sample to determine candidate sentiment polarity prior information of the unlabeled customer service text sample; For each unlabeled customer service text sample, combined with the text topology units adjacent to the unlabeled customer service text sample in the text topology structure, the candidate sentiment polarity prior information of the unlabeled customer service text sample is cyclically adjusted until the prior information of the text topology units in the text topology structure no longer changes, thereby obtaining the sentiment polarity prior information corresponding to the unlabeled customer service text sample.

3. The method according to claim 2, characterized in that The method of obtaining a semantic unity measure between the text paragraph and the customer service text sample corresponding to the text paragraph based on the pending sentiment polarity prior information set of the text paragraph and the pending sentiment polarity prior information set of the customer service text sample corresponding to the text paragraph includes: Determining the number of occurrences of the sentiment polarity prior information in the pending sentiment polarity prior information set of the text paragraph, and determining the sentiment polarity probability distribution of the text paragraph; Determine the number of occurrences of sentiment polarity prior information in the pending sentiment polarity prior information set of the customer service text sample corresponding to the text paragraph, and determine the sentiment polarity probability distribution of the customer service text sample corresponding to the text paragraph; According to the sentiment polarity probability distribution of the text paragraph and the sentiment polarity probability distribution of the customer service text sample corresponding to the text paragraph, a semantic unity measure between the text paragraph and the customer service text sample corresponding to the text paragraph is obtained.

4. The method according to claim 1, characterized in that The updating of the network weights of the basic sentiment polarity analysis network according to the semantic unity metric and the sentiment polarity prior information corresponding to the unlabeled customer service text sample to obtain a converged sentiment polarity analysis network includes: According to the sample description vector of the customer service text sample, the sentiment polarity analysis is performed on the customer service text sample to obtain a sample sentiment polarity probability distribution corresponding to the customer service text sample; According to the paragraph description vector corresponding to the text paragraph, the sentiment polarity analysis is performed on the text paragraph to obtain the paragraph sentiment polarity probability distribution corresponding to the text paragraph; Combined with the sample sentiment polarity probability distribution, the paragraph sentiment polarity probability distribution, the semantic unity measure and the sentiment polarity prior information corresponding to the unlabeled customer service text sample, the network weights of the basic sentiment polarity analysis network are updated to obtain a converged sentiment polarity analysis network.

5. The method according to claim 4, characterized in that The method combines the sample sentiment polarity probability distribution, the paragraph sentiment polarity probability distribution, the semantic unity measure and the sentiment polarity prior information corresponding to the unlabeled customer service text sample to update the network weights of the basic sentiment polarity analysis network to obtain a converged sentiment polarity analysis network, including: For each customer service text sample, based on the paragraph sentiment polarity probability distribution of each text paragraph in the customer service text sample and the semantic unity measurement corresponding to each text paragraph, the sentiment polarity prior information of the customer service text sample is optimized to obtain the optimized sentiment polarity prior information of the customer service text sample; According to the sample sentiment polarity probability distribution corresponding to each customer service text sample, the optimized sentiment polarity prior information of each customer service text sample is integrated to obtain the fusion training cost; According to the fusion training cost, the network weights of the basic sentiment polarity analysis network are updated to obtain a converged sentiment polarity analysis network.

6. The method according to claim 5, characterized in that The updating of the network weights of the basic sentiment polarity analysis network according to the fusion training cost to obtain a converged sentiment polarity analysis network includes: Obtaining a paragraph classification cost of the text paragraph according to a paragraph sentiment polarity probability distribution of the text paragraph and sentiment polarity prior information of a customer service text sample corresponding to the text paragraph; According to the fusion training cost and the paragraph classification cost, the network weights of the basic sentiment polarity analysis network are updated to obtain a converged sentiment polarity analysis network.

7. The method according to claim 6, characterized in that The updating of the network weights of the basic sentiment polarity analysis network according to the fusion training cost and the paragraph classification cost to obtain a converged sentiment polarity analysis network includes: Performing a deviation analysis between the paragraph sentiment polarity probability distribution of the text paragraph and the equal probability distribution to obtain a deviation value corresponding to the text paragraph; According to the semantic unity metric corresponding to the text paragraph, the paragraph classification cost and the deviation value are integrated to obtain the paragraph cost of the text paragraph; The network weights of the basic sentiment polarity analysis network are updated by combining the fusion training cost and the paragraph costs corresponding to each text paragraph of the customer service text sample to obtain a converged sentiment polarity analysis network.

8. A cloud server, comprising a memory and a processor, wherein the memory stores a computer program that can be run on the processor, characterized in that: When the processor executes the program, the steps in the method according to any one of claims 1 to 7 are implemented.

Images