RPA automated data entry method and computer system combined with artificial intelligence

Abstract

The present application provides an RPA automated data entry method and computer system combined with artificial intelligence, by obtaining the text semantic implicit representation and word frequency implicit representation of the RPA customer service text to be entered, and obtaining the implicit representation set of the RPA customer service text to be entered based on the text semantic implicit representation and word frequency implicit representation, the implicit representation set includes the enhanced representation vector of the RPA customer service text to be entered at one or more levels, the enhanced representation vector is obtained based on the word frequency implicit representation and the semantic representation vector contained in the text semantic implicit representation, and the RPA customer service text to be entered is identified based on the implicit representation set to obtain the abnormal conversation identification result of the RPA customer service text to be entered. Then, the semantic representation vector of the RPA customer service text to be entered is enhanced based on the word frequency implicit representation of the RPA customer service text to be entered, so that the characteristics of the local information of the RPA customer service text to be entered are amplified, and the accuracy and reliability of the abnormal conversation identification results in the text are increased.

Description

Technical Field

[0001] The present application relates to the field of data processing, and more specifically, to an RPA automated data entry method and computer system combined with artificial intelligence. Background Art

[0002] In today's digital business environment, enterprises widely use customer service systems to handle interactions with customers. With the continuous growth of business volume, manual customer service faces tremendous pressure, so RPA technology has been introduced into the customer service field to improve efficiency and reduce costs. In the process of RPA customer service text processing, accurately identifying abnormal conversations is a crucial task. Abnormal conversations may contain urgent questions, dissatisfaction with products or services, or some situations that require special handling. Traditional conversation recognition methods often have some limitations. Existing methods usually lack in-depth mining and enhanced representation of text semantics at different levels. Customer service texts often have complex semantic structures, and simple overall semantic analysis is difficult to capture key features in local information. For example, a long customer service text may contain multiple clauses, each of which may hide clues related to abnormal conversations, but traditional methods may not be able to effectively amplify the features of these local information, which in turn affects the accuracy and reliability of abnormal conversation recognition results. Summary of the invention

[0003] The purpose of the present invention is to provide an RPA automated data entry method and computer system combined with artificial intelligence. This application is implemented as follows:

[0004] In a first aspect, the present application provides an RPA automated data entry method combined with artificial intelligence, the method comprising: obtaining a text semantic implicit representation and a word frequency implicit representation of a text to be entered into an RPA customer service text; the text semantic implicit representation comprises a semantic representation vector of the RPA customer service text to be entered at a first level and a semantic representation vector of the RPA customer service text to be entered at a second level; based on the text semantic implicit representation and the word frequency implicit representation, obtaining an implicit representation set of the RPA customer service text to be entered; the implicit representation set comprises the semantic representation vector of the RPA customer service text to be entered at a second level; The semantic representation vector at the first level and the enhanced representation vectors of the RPA customer service text to be entered at x levels, the enhanced representation vector of the RPA customer service text to be entered at any level is obtained based on the semantic representation vector of the RPA customer service text to be entered at the second level and the word frequency implicit representation, x≥1; according to the implicit representation set, abnormal session recognition is performed on the RPA customer service text to be entered to obtain an abnormal session recognition result of the RPA customer service text to be entered; based on the abnormal session recognition result, the RPA customer service text to be entered is entered.

[0005] In a second aspect, the present application provides a computer system comprising: one or more processors; a memory; one or more computer programs; wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, the method as described above is implemented.

[0006] The beneficial effects of the present application include: the present application obtains the text semantic implicit representation and word frequency implicit representation of the RPA customer service text to be entered, and obtains the implicit representation set of the RPA customer service text to be entered based on the text semantic implicit representation and word frequency implicit representation, the implicit representation set includes the enhanced representation vectors of the RPA customer service text to be entered at one or more levels, the enhanced representation vectors are obtained based on the word frequency implicit representation and the semantic representation vectors contained in the text semantic implicit representation, and the RPA customer service text to be entered is identified with abnormal conversations based on the implicit representation set to obtain the abnormal conversation identification result of the RPA customer service text to be entered. Then, the semantic representation vector of the RPA customer service text to be entered is enhanced based on the word frequency implicit representation of the RPA customer service text to be entered, so that the characteristics of the local information of the RPA customer service text to be entered are amplified, and the accuracy and reliability of the abnormal conversation identification results in the text are increased. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in describing the embodiments of the present application are briefly introduced below.

[0008] Figure 1 This is a flowchart of an RPA automated data entry method combined with artificial intelligence provided in an embodiment of the present application.

[0009] Figure 2 It is a schematic diagram of the composition of a computer system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0010] The execution subject of the RPA automated data entry method combined with artificial intelligence in the embodiment of the present application is a computer system, including but not limited to a server, a personal computer, a laptop, a tablet computer, a smart phone, etc. Figure 1 As shown, the method includes:

[0011] Step S100: Obtain the text semantic implicit representation and word frequency implicit representation of the RPA customer service text to be entered; the text semantic implicit representation includes the semantic representation vector of the RPA customer service text to be entered at the first level and the semantic representation vector of the RPA customer service text to be entered at the second level.

[0012] In an embodiment of the present application, step S100 obtains the text semantic implicit representation and word frequency implicit representation of the RPA customer service text to be entered. Among them, the text semantic implicit representation covers the semantic representation vector of the RPA customer service text to be entered at the first level and the semantic representation vector at the second level. In an embodiment of the present application, different levels represent different feature extraction ranges, that is, corresponding to different scales. The computer system embeds and maps the RPA customer service text to be entered based on the first embedding mapping component. For example, assume that the RPA customer service text to be entered is "Customer feedback product B has intermittent failures." In this scenario, the first embedding mapping component may use a pre-trained word vector model, such as Word2Vec. Assume that the word vector dimension of the Word2Vec model is 300 dimensions. For each word in this customer service text, the computer system will look up the corresponding 300-dimensional vector representation in the Word2Vec model. Then, the vectors of these words are combined through specific combination rules (such as weighted average, etc.) to obtain the semantic representation vector of the RPA customer service text to be entered at the first level. Assume that the vector obtained after calculation is [0.1, 0.2, -0.3, …, 0.5] (this is just an example, the actual dimension is 300 dimensions). Then, the computer system embeds and maps the semantic representation vector at the first level based on the second embedding mapping component, so as to obtain the semantic representation vector at the second level of the RPA customer service text to be entered. Continuing with the above customer service text as an example, the second embedding mapping component may be a mapping layer based on a neural network. Assume that this neural network mapping layer contains 100 neurons. The semantic representation vector at the first level is used as input and is transformed by the nonlinear transformation of the neural network (for example, using the activation function ReLU(x)=max(0,x)). After calculation by this neural network layer, the semantic representation vector at the second level of the RPA customer service text to be entered is obtained. Assume that the obtained vector is [-0.1, 0.3, 0.4, …, -0.2] (also just an example, the actual dimension depends on the output setting of the neural network). The semantic representation vectors at these two levels together constitute the text semantic implicit representation of the RPA customer service text to be entered.

[0013] In addition, the computer system performs word frequency statistics on the RPA customer service text to be entered through the word frequency channel. Taking "The customer inquires about the function, price, and whether there are any promotional activities of product A" as an example, the computer system will count the number of times each word appears in this text. For example, "product A" appears once, "function" appears once, "price" appears once, "promotional activities" appears once, "customer" appears once, "inquiry" appears once, and "whether there are" appears once as a phrase. Then, the word frequency statistics results are quantized based on the word frequency table. Assume that the quantization rule set in the word frequency table is: words that appear once are quantized to 1, words that appear twice are quantized to 2, and so on. After quantization, these quantized values ​​constitute the word frequency implicit representation of the RPA customer service text to be entered, such as [1, 1, 1, 1, 1, 1, 1] (corresponding to the quantization values ​​of each of the above words or phrases). In the entire step S100, the computer system accurately obtains the text semantic implicit representation and word frequency implicit representation of the RPA customer service text to be entered through these operations. These two representations lay the foundation for subsequent steps, such as obtaining a set of implicit representations of the RPA customer service text to be entered based on these representations in step S200. By obtaining semantic representation vectors and counting word frequencies in this hierarchical manner, the computer system can more comprehensively and deeply understand the semantic information and word frequency distribution characteristics of the RPA customer service text to be entered.

[0014] Step S200: Obtain an implicit representation set of the RPA customer service text to be entered based on the implicit representation of text semantics and the implicit representation of word frequency; the implicit representation set includes the semantic representation vector of the RPA customer service text to be entered at the first level and the enhanced representation vectors of the RPA customer service text to be entered at x levels, and the enhanced representation vector of the RPA customer service text to be entered at any level is obtained based on the semantic representation vector of the RPA customer service text to be entered at the second level and the implicit representation of word frequency, x≥1.

[0015] In an embodiment of the present application, step S200 obtains an implicit representation set of the text to be entered into the RPA customer service text based on the text semantic implicit representation and word frequency implicit representation obtained in step S100.

[0016] First, the computer system integrates the semantic representation vector and word frequency implicit representation of the RPA customer service text to be entered at the second level to obtain an enhanced word frequency implicit representation. For example, assuming that the RPA customer service text to be entered is "User feedback that compatibility issues occurred after the software update of product C", the semantic representation vector at the second level obtained in step S100 is [0.2, -0.1, 0.3, ..., -0.4] (to simplify the example, assume that the dimension is 5 dimensions), and the word frequency implicit representation is [1, 1,1, 1, 1] (corresponding to the quantized values ​​of different words or phrases).

[0017] Computer systems can be integrated using specific neural network structures. For example, a small neural network with multiple fully connected layers is constructed. For the position focus strategy, it can be implemented based on the position attention mechanism, by weighting the position relationship of the elements in the semantic representation vector and the word frequency implicit representation. Assuming the weight matrix is ​​W1 (dimension is 5×5), the calculation method is to multiply the semantic representation vector with the weight matrix and add the word frequency implicit representation to obtain the first attention implicit representation: V1=[0.2, -0.1, 0.3, -0.4]*W1 + [1, 1, 1, 1, 1]. For the multi-head focus strategy, assuming that three heads are used, each head has a different weight matrix (such as W2, W3, W4, all with dimensions of 5×5), and three intermediate results are calculated separately, and then these three results are spliced ​​together to obtain the second attention implicit representation. Finally, the fusion result of the first attention implicit representation and the second attention implicit representation is processed by a feature extraction component (such as a convolution layer). Assuming that the convolution kernel is K (dimension is 3×3), the fusion result is convolved to obtain the enhanced word frequency implicit representation. Then, the computer system integrates the semantic representation vector of the RPA customer service text to be entered at the second level with the enhanced word frequency implicit representation to obtain the enhanced representation vector of the RPA customer service text to be entered at the second level. Continuing with the above example, assume that the enhanced word frequency implicit representation obtained after the above operation is [0.5, 0.3, -0.2, 0.1, 0.4]. The computer system uses a target feature extraction network (such as a simple linear layer) to extract features from the enhanced word frequency implicit representation. Assuming that the weight matrix of the linear layer is W5 (dimension is 5×3), the feature extraction result is [0.2, 0.1, -0.1]. Then, the feature extraction result is summed element by element with the semantic representation vector at the second level, and the enhanced representation vector of the text to be entered into the RPA customer service at the second level is obtained as [0.4, 0.0, 0.2, …, -0.3].

[0018] If x>1, the computer system repeatedly embeds and maps (i.e., encodes) the enhanced representation vector of the RPA customer service text to be entered at the second level based on y embedding mapping components to obtain the enhanced representation vector of the RPA customer service text to be entered at y levels, y≥1. For example, assuming y = 2, for the enhanced representation vector [0.4, 0.0,0.2, …, -0.3] at the second level, the first embedding mapping component is a neural network with a hidden layer containing 5 neurons, which takes the enhanced representation vector as input, and obtains the enhanced representation vector at the third level through calculation by the neural network (assuming that the activation function is the Sigmoid function, Sigmoid(x)=1 / (1+ e^(-x))). The second embedding mapping component is a neural network with a hidden layer containing 8 neurons, which takes the enhanced representation vector at the third level as input, and obtains the enhanced representation vector at the fourth level through calculation.

[0019] Finally, the computer system obtains the implicit representation set of the RPA customer service text to be entered based on the semantic representation vector of the RPA customer service text to be entered at the first level and the enhanced representation vector of the RPA customer service text to be entered at the second level (if there are multiple enhanced representation vectors at the level), For example, the semantic representation vector at the first level is [0.1, 0.2, -0.3, …, 0.5], and the enhanced representation vector at the second level is [0.4, 0.0, 0.2, …, -0.3]. If there are enhanced representation vectors at the third level and the fourth level that are [0.3, 0.1, -0.2, …,0.4] and [-0.1, 0.2, 0.1, …, -0.2] respectively, then these vectors together constitute the implicit representation set of the RPA customer service text to be entered. In the entire step S200, the computer system, through a series of operations, starts from the original text semantic implicit representation and word frequency implicit representation, and constructs an implicit representation set of the RPA customer service text to be entered through integration, feature extraction, embedding mapping and other operations. This implicit representation set contains semantic related vectors at different levels, which provides a comprehensive and rich data foundation for subsequent steps such as abnormal session identification of the RPA customer service text to be entered in step S300. In this way, the computer system can more deeply mine the semantic information in the RPA customer service text to be entered, and improve the accuracy and reliability of the automated data entry process.

[0020] Step S300: Based on the implicit representation set, abnormal session identification is performed on the RPA customer service text to be entered, and an abnormal session identification result of the RPA customer service text to be entered is obtained.

[0021] In an embodiment of the present application, step S300 performs abnormal conversation identification on the RPA customer service text to be entered based on the implicit representation set obtained in step S200, thereby obtaining the abnormal conversation identification result of the RPA customer service text to be entered. A classifier is a machine learning model, such as a support vector machine (SVM) classifier. Assume that the implicit representation set obtained in step S200 includes the semantic representation vector of the RPA customer service text to be entered at the first level as [0.1, 0.2, -0.3, 0.4], and the enhanced representation vector at the second level as [-0.1, 0.3, 0.2, -0.2] (this is a simplified example, and the actual vector dimension may be higher). If there are other enhanced representation vectors at other levels, they are also included in this set.

[0022] For the SVM classifier, its working principle is to find a hyperplane to separate data of different categories. In this scenario, the data is the vector in the implicit representation set, and the categories are normal conversations and abnormal conversations. The computer system first needs to preprocess the vectors in the implicit representation set to make them meet the input requirements of the SVM classifier. For example, multiple vectors are concatenated into a high-dimensional vector, assuming that [0.1, 0.2, -0.3, 0.4, -0.1, 0.3, 0.2, -0.2] is obtained. The SVM classifier performs classification by learning a decision function. Assume that the decision function is f(x)=w·x + b, where x is the input vector (that is, the implicit representation set vector after preprocessing), w is the weight vector, and b is the bias term. In the training phase, the computer system determines the values ​​of w and b through the implicit representation set vectors corresponding to a large number of labeled (normal or abnormal) RPA customer service texts. In the prediction phase, when the implicit representation set vector of the RPA customer service text to be entered is input, the value of f(x) is calculated. If f(x)>0, it is determined to be a normal session; if f(x)≤0, it is determined to be an abnormal session.

[0023] Step S400: Enter the RPA customer service text to be entered based on the abnormal session identification result.

[0024] In the embodiment of the present application, step S400 performs an entry operation on the RPA customer service text to be entered based on the abnormal session recognition result obtained in step S300.

[0025] After the computer system obtains the abnormal conversation recognition results of the RPA customer service text to be entered, it will adopt corresponding entry strategies according to different situations in the results. Assuming that the RPA customer service text to be entered is "User feedback that unknown noise occurs during the use of product E", its abnormal conversation recognition results indicate that there is no abnormality in the conversation. At this time, the computer system will enter the entire text completely and accurately into the corresponding database or system. For example, if the target database is a customer feedback management system, the computer system will accurately correspond each element in the text to the corresponding field in the database according to the established data format, such as "user" corresponds to the feedback field, "product E" corresponds to the product name field, and "unknown noise occurs during use" corresponds to the feedback content field. However, if the abnormal conversation recognition results show that there is an abnormal situation, the computer system will perform special processing before entering. For example, suppose the RPA customer service text to be entered is "The customer claims that the maintenance date of product F has passed, but no notification has been received, and requires immediate resolution", and the abnormal conversation recognition results indicate that the statement "the maintenance date has passed but no notification has been received" is abnormal. The computer system may mark or adjust the abnormal part before entering. For example, when entering information into the customer feedback management system, the computer system will add a special mark next to the content "The maintenance date has passed but no notification has been received" in the "Feedback Content" field, indicating that this is an abnormal point so that subsequent staff can quickly identify and handle it.

[0026] As an implementation manner, step S100, obtaining the text semantic implicit representation and word frequency implicit representation of the text to be entered into the RPA customer service text, may include:

[0027] Step S110: embedding and mapping the RPA customer service text to be input based on the first embedding mapping component to obtain a semantic representation vector of the RPA customer service text to be input at the first level;

[0028] Step S120: embedding and mapping the semantic representation vector of the RPA customer service text to be entered at the first level based on the second embedding mapping component to obtain the semantic representation vector of the RPA customer service text to be entered at the second level;

[0029] Step S130: using the semantic representation vector of the RPA customer service text to be entered at the first level and the semantic representation vector of the RPA customer service text to be entered at the second level as the text semantic implicit representation of the RPA customer service text to be entered;

[0030] Step S140: performing word frequency statistics on the text to be input into the RPA customer service through a word frequency channel to obtain a word frequency statistical result of the text to be input into the RPA customer service;

[0031] Step S150: quantify the word frequency statistics of the RPA customer service text to be entered based on the word frequency table to obtain an implicit representation of the word frequency of the RPA customer service text to be entered.

[0032] Step S110 converts the RPA customer service text to be input into a semantic representation vector at the first level. The first embedding mapping component is a key element to achieve this conversion.

[0033] For example, suppose the RPA customer service text to be entered is "The customer inquired about the new features of product G and its compatibility with other products." In this scenario, the first embedding mapping component may use a pre-trained word vector model, such as the GloVe (Global Vectors for Word Representation) model. The GloVe model is a word vector model based on global word frequency statistics. It generates a fixed-dimensional vector representation for each word by training a large-scale corpus. Assume that the word vector dimension generated by the GloVe model is 300 dimensions. For each word in the above customer service text, the computer system will look up the corresponding 300-dimensional vector in the GloVe model. For example, the word "customer" has a corresponding 300-dimensional vector in the GloVe model, "inquiry" also has its corresponding vector, and so on. Then, the computer system combines the vectors of these words through specific combination rules to obtain the semantic representation vector of the RPA customer service text to be entered at the first level. A common combination rule is the weighted average method.

[0034] Assume that in this customer service text, the vector of "customer" is , and the vector of “query” is , and so on. If a simple average weighting is used (each word has the same weight), and there are n words in the text, then the semantic representation vector at the first level is The calculation method is:

[0035] ;

[0036] Assume that the calculated vector is (This is just an example, the actual dimension is 300.) In this way, the computer system uses the first embedding mapping component and predefined calculation rules to convert the RPA customer service text to be entered into a semantic representation vector at the first level, which will serve as the basis for subsequent processing.

[0037] After completing step S110, the computer system enters step S120. The purpose of this step is to further convert the semantic representation vector at the second level of the to-be-entered RPA customer service text through the second embedding mapping component based on the semantic representation vector at the first level that has been obtained.

[0038] Continuing with the above RPA customer service text "Customer inquires about the new features of product G and its compatibility with other products" as an example, the semantic representation vector at the first level obtained in step S110 is The second embedding mapping component may be a neural network based mapping layer, such as a simple multilayer perceptron (MLP).

[0039] Assume that this MLP has an input layer, a hidden layer, and an output layer. The number of neurons in the input layer is the same as the dimension of the semantic representation vector in the first layer, that is, 300 neurons (consistent with the dimension of the GloVe word vector assumed above). The hidden layer can be set to 100 neurons, and the number of neurons in the output layer is set according to the needs, assuming it is 200-dimensional. This 200-dimensional vector will be used as the semantic representation vector in the second layer.

[0040] When embedding, the semantic representation vector at the first level is used as the input of the MLP. The connection between each neuron has a corresponding weight. For the connection from the input layer to the hidden layer, the weight matrix is ​​assumed to be W1 (dimension is 300×100) and the bias vector is b1 (dimension is 100). The activation function of the hidden layer neuron can be the ReLU (Rectified Linear Unit) function, that is, . Then the output h of the hidden layer neuron is calculated as follows:

[0041] ;

[0042] Next, from the hidden layer to the output layer, assuming that the weight matrix is ​​W2 (dimension is 100×200) and the bias vector is b2 (dimension is 200), the semantic representation vector at the second level is Calculated as: ; Assume that the semantic representation vector at the second level obtained by calculation is (This is just an example, the actual dimension is 200.) Through this process, the computer system uses the second embedding mapping component to further embed the semantic representation vector at the first level to obtain the semantic representation vector at the second level, which helps to mine deeper semantic information of the text.

[0043] In step S130, the computer system combines the two vectors obtained in steps S110 and S120, namely, the semantic representation vector of the RPA customer service text to be entered at the first level and the semantic representation vector at the second level, as the text semantic implicit representation of the RPA customer service text to be entered.

[0044] Taking the above example, the semantic representation vector at the first level is , the semantic representation vector at the second level is , these two vectors together constitute the implicit representation of text semantics. This combination can describe the semantics of the RPA customer service text to be entered from different levels. In the subsequent processing steps, this implicit representation of text semantics will be used as important input data for further analysis and processing. For example, in operations such as abnormal session identification, it can provide the computer system with more comprehensive and richer semantic information to better understand the content of the RPA customer service text to be entered.

[0045] When the computer system executes step S140, it starts to perform word frequency statistics on the RPA customer service text to be entered. Still taking "Customer inquires about the new features of product G and its compatibility with other products" as an example, the computer system performs word frequency statistics through the word frequency channel. In this text, "customer" appears once, "inquiry" appears once, "product G" appears once, "new features" appears once, "other products" appears once, and "compatibility" appears once. The computer system records the number of occurrences of each word (or phrase, if "product G", "new features", "other products", etc. are regarded as phrases) to obtain the word frequency statistics results. This result can be expressed as an array, for example: ; Each element in this array is a tuple, the first element of the tuple is a word or phrase, and the second element is the number of times it appears in the text. In this way, the computer system can accurately count the frequency of each word or phrase in the text to be entered into the RPA customer service, providing a basis for the next step of quantitative processing.

[0046] In step S150, the computer system quantifies the word frequency statistics obtained in step S140 based on the word frequency table, thereby obtaining an implicit representation of the word frequency to be entered into the RPA customer service text.

[0047] Assume that the word frequency table sets the following quantization rules: words or phrases that appear once are quantized to 1, words or phrases that appear twice are quantized to 2, words or phrases that appear three times are quantized to 3, and so on. For the word frequency statistics results [("customer", 1), ("inquiry", 1), ("product G", 1), ("new function", 1), ("other products", 1), ("compatibility", 1)] obtained in the previous step S140, after quantization processing, the word frequency implicit representation obtained is an array: [1, 1, 1, 1, 1, 1], and each element in this array corresponds to the quantization value of each word or phrase in the original word frequency statistics results. For example, the first element 1 corresponds to the quantization value of "customer", the second element 1 corresponds to the quantization value of "inquiry", and so on. Through this quantization processing, the computer system converts the word frequency statistics results into a form that is more suitable for subsequent processing, namely, the word frequency implicit representation. In the entire RPA automated data entry method combined with artificial intelligence, the word frequency implicit representation, together with the previously obtained text semantic implicit representation, provides the necessary data basis for the construction of the implicit representation set in subsequent steps such as step S200, which helps to further analyze and process the RPA customer service text to be entered.

[0048] In this implementation of the entire step S100, the computer system gradually obtains the implicit semantic representation and word frequency implicit representation of the text from the RPA customer service text to be entered through the above five steps. These representations lay a solid foundation for data processing, analysis and final automated data entry operations in subsequent steps, enabling the computer system to better understand the RPA customer service text to be entered and perform subsequent operations based on these understandings. At the same time, the specific models (such as GloVe model, MLP, etc.) and calculation methods (such as weighted average, forward propagation calculation of neural network, etc.) involved in each step are operated according to predetermined rules and algorithms to ensure the accuracy and effectiveness of the entire process.

[0049] As an implementation mode, in step S200, obtaining an implicit representation set of the text to be input into the RPA customer service text according to the implicit representation of text semantics and the implicit representation of word frequency may include:

[0050] Step S210: Integrate the semantic representation vector and word frequency implicit representation of the RPA customer service text to be entered at the second level to obtain an enhanced word frequency implicit representation;

[0051] Step S220: Integrate the semantic representation vector of the RPA customer service text to be entered at the second level with the enhanced word frequency implicit representation to obtain the enhanced representation vector of the RPA customer service text to be entered at the second level;

[0052] Step S230: Obtain an implicit representation set of the RPA customer service text to be entered based on the semantic representation vector of the RPA customer service text to be entered at the first level and the enhanced representation vector of the RPA customer service text to be entered at the second level.

[0053] Step S210 combines the semantic representation vector of the text to be entered into the RPA customer service at the second level with the implicit representation of word frequency through a specific integration operation, thereby obtaining an enhanced implicit representation of word frequency. This process helps to explore the potential connection between text semantics and word frequency, and provides a basis for subsequent representation vector enhancement.

[0054] For example, suppose the RPA customer service text to be entered is "User feedback that product H runs unstably in a specific environment and an error message appears." In step S100, the semantic representation vector and word frequency implicit representation of the text at the second level have been obtained. Assume that the semantic representation vector at the second level is (To simplify the example, assume that the dimension is 5), the word frequency is implicitly represented as [1, 1, 1, 1, 1] (corresponding to the quantized values ​​of different words or phrases). The computer system may use a neural network structure based on the attention mechanism for integration. The attention mechanism can automatically learn the association weights between different elements to integrate information more effectively. First, for the position focus strategy, the computer system performs weighted calculations on the element position relationship in the semantic representation vector and the word frequency implicit representation. Assume that a position weight matrix is ​​defined (dimension is 5×5), the calculation method is to multiply the semantic representation vector with the position weight matrix and add the word frequency implicit representation to get the first attention implicit representation : ; Assuming that after calculation .

[0055] Next, for the multi-head focus strategy, assume that three heads are used, each with a different weight matrix (such as , with dimensions of 5×5). For each head, the semantic representation vector is multiplied by the corresponding weight matrix and then added to the implicit representation of word frequency to obtain three intermediate results. For example, for the first head: ;

[0056] Assume that = [1.1, 1.0, 1.2, 0.7, 1.0]. Similarly, we can calculate and , and then concatenate these three results in a certain dimension (such as column concatenation) to obtain the second attention implicit representation Assuming that the spliced is [1.1, 1.0, 1.2, 0.7, 1.0, 1.3, 1.1, 1.4, 0.8, 1.2, 1.2, 1.0,1.3, 0.9, 1.1].

[0057] Finally, the first attention implicit representation V_{1} and the second attention implicit representation V_{1} are extracted through a feature extraction component (such as a convolutional layer). Assume that the convolution kernel is K (dimension is 3×3), and process the fusion result (assuming that and After concatenating by row, a vector with a dimension of 1×20 is obtained (here for simplification assumptions) for convolution operation. The convolution operation is calculated as follows: for each position where the convolution kernel slides on the input vector, the sum of the products of the convolution kernel elements and the corresponding input vector elements is calculated. Assume that after the convolution operation, the enhanced word frequency implicit representation is obtained as [0.5, 0.3, -0.2, 0.1, 0.4]. In step S220, the computer system integrates the semantic representation vector of the RPA customer service text to be entered at the second level with the enhanced word frequency implicit representation obtained in step S210 to obtain the enhanced representation vector of the text at the second level. This step further integrates the semantic information and the enhanced word frequency information, making the representation vector more representative.

[0058] Continuing with the above example, assume that the implicit representation of the frequency of the enhanced words obtained after the above operation is [0.5, 0.3, -0.2, 0.1, 0.4], and the semantic representation vector of the text to be entered into the RPA customer service at the second level is [0.2, -0.1, 0.3, , -0.4].

[0059] The computer system uses a target feature extraction network (e.g., a simple linear layer) to extract features from the implicit representation of the reinforced word frequency. Assume that the weight matrix of this linear layer is (dimension is 5×3), the bias vector is (The dimension is 3). Then the feature extraction result The calculation is as follows:

[0060] ;

[0061] Assuming that the calculation = [0.2, 0.1, -0.1].

[0062] Then, the feature extraction result is summed element by element with the semantic representation vector at the second level to obtain the enhanced representation vector at the second level of the text to be entered into the RPA customer service. : .

[0063] In step S230, the computer system constructs an implicit representation set of the RPA customer service text to be entered based on the semantic representation vector of the RPA customer service text to be entered at the first level and the enhanced representation vector at the second level. This implicit representation set contains semantic information obtained from different levels, providing a comprehensive data foundation for subsequent operations (such as abnormal session identification).

[0064] For example, suppose the semantic representation vector of the RPA customer service text to be entered at the first level is [0.1, 0.2, -0.3, , 0.5], and the enhanced representation vector at the second level is [0.4, 0.0, 0.2, , -0.5]. Then the computer system combines these two vectors to form an implicit representation set.

[0065] If step S200 involves the calculation of enhanced representation vectors of more levels (for example, the enhanced representation vectors of the second level are repeatedly embedded and mapped based on y embedding mapping components to obtain enhanced representation vectors of other levels), these vectors will also be included in the implicit representation set. Assume that after further calculation, the enhanced representation vector of the third level is [0.3, 0.1, -0.2, , 0.4], then the implicit representation set will contain these three vectors: Implicit representation set = {[0.1, 0.2, -0.3, , 0.5], [0.4, 0.0, 0.2, , -0.5], [0.3,0.1, -0.2, , 0.4]}.

[0066] This implicit representation set comprehensively reflects the semantic features of the RPA customer service text to be entered at different levels, and will serve as the input for subsequent steps (such as abnormal session identification in step S300). In this implementation of the entire step S200, the computer system obtains the implicit representation set of the RPA customer service text to be entered through the above three steps, starting from the existing text semantic implicit representation and word frequency implicit representation, and after a series of integration and construction operations.

[0067] As an implementation method, the method provided in the embodiment of the present application may further include the following steps:

[0068] Step S201: Repeat embedding and mapping the enhanced representation vector of the RPA customer service text to be entered at the second level based on y embedding mapping components to obtain the enhanced representation vector of the RPA customer service text to be entered at y levels, y ≥ 1. Based on this, step S230, according to the semantic representation vector of the RPA customer service text to be entered at the first level and the enhanced representation vector of the RPA customer service text to be entered at the second level, obtain the implicit representation set of the RPA customer service text to be entered, including: the semantic representation vector of the RPA customer service text to be entered at the first level, the enhanced representation vector of the RPA customer service text to be entered at the second level, and the enhanced representation vector of the RPA customer service text to be entered at y levels are aggregated to obtain the implicit representation set of the RPA customer service text to be entered.

[0069] Step S201 repeatedly embeds and maps the enhanced representation vector of the RPA customer service text to be entered at the second level through multiple embedding mapping components, thereby obtaining enhanced representation vectors at y levels. This operation helps to further mine the semantic information of the text and characterize the text from different levels. For example, suppose the RPA customer service text to be entered is "Customer feedback shows that the performance of product I has declined after the update", and the enhanced representation vector of the text at the second level has been obtained in the previous step, assuming it is [0.3, -0.1, 0.2, , -0.3]. Assume y = 2, that is, the computer system will use 2 embedding mapping components for repeated embedding mapping.

[0070] The first embedding mapping component can be a neural network structure based on a multi-layer perceptron (MLP). Assume that this MLP has an input layer, a hidden layer, and an output layer. The number of neurons in the input layer is the same as the dimension of the reinforcement representation vector in the second layer, that is, 5 neurons. The hidden layer is set to 3 neurons, and the number of neurons in the output layer is set according to the requirements, assuming it is 4-dimensional. This 4-dimensional vector will be used as the reinforcement representation vector in the third layer. When performing embedding mapping, the reinforcement representation vector in the second layer is used as the input of the MLP. For the connection from the input layer to the hidden layer, assume that the weight matrix is (dimension is 5×3), the bias vector is (The dimension is 3D), the activation function of the hidden layer neurons uses the ReLU (Rectified Linear Unit) function, that is, ReLU(x) = max(0, x). Then the output h of the hidden layer neurons is calculated as follows:

[0071] ;

[0072] Next, from the hidden layer to the output layer, assuming the weight matrix is (dimension is 3×4), the bias vector is (The dimension is 4), then the enhanced representation vector at the third level is Calculated as: ; Assume that the calculated enhanced representation vector at the third level is [0.2, 0.1, -0.1, 0.4].

[0073] The second embedding mapping component is also an MLP with a similar structure, but the parameters may be different. For example, the number of neurons in the input layer is 4 (the same as the dimension of the reinforcement representation vector at the third level), the hidden layer is set to 2 neurons, and the output layer is 3-dimensional. According to the above similar calculation process, the reinforcement representation vector at the third level is used as input, and the reinforcement representation vector at the fourth level is calculated by this MLP, which is assumed to be [-0.1, 0.2, 0.1].

[0074] In this way, the computer system repeatedly embeds and maps the enhanced representation vector of the RPA customer service text to be entered at the second level based on y embedding mapping components (here y = 2), and obtains the enhanced representation vectors at y levels (here the third and fourth levels).

[0075] In step S230, the computer system collects the representation vectors of different levels obtained previously and constructs an implicit representation set of the text to be entered into the RPA customer service. This implicit representation set integrates the semantic information of the text at multiple levels, providing a rich data foundation for subsequent processing (such as abnormal session identification, etc.). Continuing with the above example, assume that the semantic representation vector of the RPA customer service text to be entered at the first level is [0.1, 0.2, -0.3, , 0.5], and the enhanced representation vector at the second level is [0.3, -0.1, 0.2, , -0.3], the enhanced representation vector at the third level is [0.2, 0.1, -0.1, 0.4], and the enhanced representation vector at the fourth level is [-0.1, 0.2, 0.1]. The computer system aggregates these vectors to obtain the implicit representation set to be entered into the RPA customer service text: Implicit representation set = {[0.1, 0.2, -0.3, , 0.5], [0.3, -0.1, 0.2, , -0.3], [0.2, 0.1, -0.1, 0.4],[ -0.1, 0.2, 0.1]}.

[0076] As an implementation mode, step S210 integrates the semantic representation vector and the word frequency implicit representation of the RPA customer service text to be input at the second level to obtain the enhanced word frequency implicit representation, which may include:

[0077] Step S211: Processing the semantic representation vector and word frequency implicit representation of the RPA customer service text to be input at the second level based on the position focus strategy to obtain the first attention implicit representation;

[0078] Step S212: Based on the multi-head focus strategy, the semantic representation vector and word frequency implicit representation of the RPA customer service text to be input at the second level are processed to obtain the second attention implicit representation;

[0079] Step S213: Process the fusion result of the first attention implicit representation and the second attention implicit representation based on the feature extraction component to obtain the enhanced word frequency implicit representation.

[0080] When executing step S211, the computer system uses a position focus strategy to process the semantic representation vector and word frequency implicit representation of the RPA customer service text to be entered at the second level, aiming to explore the potential connection between the two based on the position relationship, thereby obtaining the first attention implicit representation.

[0081] For example, suppose the RPA customer service text to be entered is "Users reported that product J has scratches on its appearance and the packaging is damaged." In the previous step, the semantic representation vector of the text at the second level has been obtained, which is assumed to be [0.2, -0.1, 0.3, 0.4, -0.2] (this is a simplified example, assuming the dimension is 5), and the word frequency is implicitly represented as [1, 1, 1, 1, 1] (corresponding to the quantized values ​​of different words or phrases). The position focus strategy adopted by the computer system may be based on a predefined position weight matrix. Assume that this position weight matrix is , with a dimension of 5×5. The position focus strategy is calculated by multiplying the semantic representation vector with the position weight matrix and adding the word frequency implicit representation to obtain the first attention implicit representation The specific calculation process is as follows: ; Assume the position weight matrix for: ;

[0082] but: ;

[0083] Add the word frequency implicit representation [1, 1, 1, 1, 1] to get the first attention implicit representation = [1.04,1.025, 1.04, 1.05, 1.02].

[0084] In step S212, the computer system uses a multi-head focusing strategy to further process the semantic representation vector and word frequency implicit representation of the text to be entered into the RPA customer service at the second level to obtain the second attention implicit representation. The multi-head focusing strategy can process data from multiple "heads" (i.e., different perspectives or calculation paths) to more comprehensively capture the information in the data. Continuing with the above example, assume that three heads are used for the multi-head focusing strategy. For each head, there is a different weight matrix.

[0085] Let the weight matrix of the first head be , with a dimension of 5×5, the weight matrix of the second head is , with a dimension of 5×5, the weight matrix of the third head is , with a dimension of 5×5.

[0086] For the first head, the calculation method is to add the semantic representation vector to Multiply and add the implicit representation of word frequency to get the intermediate result Assumptions for: ;

[0087] but: = ;

[0088] Similarly, calculate and , assuming = [1.06, 1.04, 1.07, 1.08, 1.01], = [1.07, 1.03, 1.08, 1.09, 1.02].

[0089] Then concatenate these three results in some way (e.g., concatenate by column) to get the second attention implicit representation Assuming that the spliced is [1.08, 1.06, 1.07, 1.05, 1.04, 1.09, 1.07, 1.08, 1.1,1.08, 0.99, 1.01, 1.02].

[0090] In step S213, the computer system processes the fusion result of the first attention implicit representation and the second attention implicit representation using a feature extraction component to obtain a reinforced word frequency implicit representation. The feature extraction component helps to extract key features from the fused complex information and further optimize the representation. =[1.04, 1.025, 1.04, 1.05, 1.02] and the second attention implicit representation = [1.08, 1.06, 1.07,1.05, 1.04, 1.09, 1.07, 1.08, 1.1, 1.08, 0.99, 1.01, 1.02] for fusion. A simple way to fuse them is to concatenate them row by row to get a fused vector Assuming that the spliced is [1.04, 1.025, 1.04, 1.05, 1.02, 1.08, 1.06, 1.07, 1.05, 1.04, 1.09, 1.07, 1.08, 1.1, 1.08, 0.99, 1.01, 1.02]. The feature extraction component can be a convolution layer. Assume that the convolution kernel is K and the dimension is 3×3. For the fusion vector Perform a convolution operation. The convolution operation is calculated by calculating the sum of the products of the convolution kernel elements and the corresponding input vector elements for each position where the convolution kernel slides on the input vector.

[0091] For example, suppose the convolution kernel K is: ; From the fusion vector Starting from the first position of , take a 3×3 subvector (here, since the dimension is 18, it is assumed that appropriate zero padding is performed to meet the convolution operation requirements). Assuming that the subvector taken is [1.04, 1.025, 1.04, 1.08, 1.06, 1.07, 1.04, 1.02,1.08], the first element of the convolution result is calculated as: ; In this way, the convolution kernel is slid on the fusion vector to calculate and obtain the implicit representation of enhanced word frequency. Assume that the implicit representation of enhanced word frequency obtained after the convolution operation is [0.5, 0.3, -0.2, 0.1, 0.4] (this is a simplified example, and the dimension matches the dimension of the calculation result assumed above).

[0092] In this implementation of the entire step S210, the computer system uses the position focus strategy, multi-head focus strategy and feature extraction component through the above three steps, starting from the semantic representation vector and word frequency implicit representation of the to-be-entered RPA customer service text at the second level, and after a series of complex calculations and processing, obtains an enhanced word frequency implicit representation.

[0093] As an implementation mode, step S220, integrating the semantic representation vector of the RPA customer service text to be entered at the second level with the enhanced word frequency implicit representation to obtain the enhanced representation vector of the RPA customer service text to be entered at the second level, may include:

[0094] Step S221: extracting features of the enhanced word frequency implicit representation based on the target feature extraction network to obtain a feature extraction result of the enhanced word frequency implicit representation;

[0095] Step S222: sum the feature extraction result implicitly represented by the enhanced word frequency and the semantic representation vector of the RPA customer service text to be entered at the second level element by element to obtain the enhanced representation vector of the RPA customer service text to be entered at the second level.

[0096] When the computer system executes step S221, it uses the target feature extraction network to perform feature extraction operations on the implicit representation of the reinforcement word frequency. The role of the target feature extraction network is to mine more representative and discriminative features from the implicit representation of the reinforcement word frequency for subsequent integration with the semantic representation vector. For example, assuming that the RPA customer service text to be entered is "Customer feedback shows that the battery life of product K is much lower than the advertised value", the implicit representation of the reinforcement word frequency obtained in the previous step is [0.5, 0.3, -0.2, 0.1, 0.4] (to simplify the example, assume that the dimension is 5). The target feature extraction network can be a simple linear layer neural network structure. Suppose the weight matrix of this linear layer is (dimension is 5×3), the bias vector is (The dimension is 3.) For the enhanced word frequency implicit representation [0.5, 0.3, -0.2, 0.1, 0.4], the feature extraction results The calculation of is as follows: ; Assume the weight matrix for: ; Bias vector = [0.1, 0.2, 0.3]; then: .

[0097] The feature extraction result of strengthening the implicit representation of word frequency is [0.24, 0.51, 0.51].

[0098] In step S222, the computer system performs element-by-element summation of the feature extraction result of the enhanced word frequency implicit representation obtained in step S221 and the semantic representation vector of the RPA customer service text to be entered at the second level, thereby obtaining the enhanced representation vector of the text at the second level. This step effectively incorporates the feature information related to word frequency into the semantic representation vector, further enriching the semantic connotation of the vector.

[0099] Continuing with the above example, assume that the semantic representation vector of the RPA customer service text to be entered at the second level is [0.2, -0.1, 0.3,…, -0.4] (the dimension here is consistent with the previous assumption, which is 5 dimensions, and the simplified representation only lists some elements).

[0100] The feature extraction result [0.24, 0.51, 0.51] (here, in order to perform element-by-element summation with the semantic representation vector, it can be regarded as [0.24, 0.51, 0.51, 0, 0], and the dimension is supplemented to match) of the implicit representation of the enhanced word frequency is summed element by element with the semantic representation vector [0.2, -0.1, 0.3, -0.4, 0]: That is, the enhanced representation vector of the text to be entered into the RPA customer service at the second level is [0.44, 0.41, 0.81, -0.4, 0].

[0101] In this implementation of the entire step S220, the computer system first uses the target feature extraction network to extract features from the reinforced word frequency implicit representation through the above two steps, and then sums the extraction result with the semantic representation vector element by element.

[0102] As an implementation mode, step S300, based on the implicit representation set, performs abnormal session identification on the RPA customer service text to be entered to obtain the abnormal session identification result of the RPA customer service text to be entered, which may include: classifying the implicit representation set based on the classifier to obtain the abnormal session identification result of the RPA customer service text to be entered; the abnormal session identification result includes the abnormal session identification result of each text segmentation in the RPA customer service text to be entered.

[0103] Assume that the RPA customer service text to be entered is "Users complained that product L had serious quality problems and was damaged shortly after use". In the previous steps, the implicit representation set of the text has been obtained, for example, the semantic representation vector contained in the first level is [0.1, 0.2, -0.3, …, 0.5], the reinforcement representation vector in the second level is [0.4, 0.0,0.2, …, -0.5], and related vectors in other levels (if any) that may exist. A classifier is a machine learning model, and a decision tree classifier is taken as an example here. A decision tree classifier builds a tree structure by learning the features and corresponding category labels in a data set to classify new data. For the vectors in the implicit representation set, the computer system uses them as input features of the decision tree classifier. Each internal node of a decision tree represents a test on an attribute, the branch is the test output, and the leaf node is a category or value. For example, a decision tree may first check whether an element in the semantic representation vector at the first level (such as the first element 0.1) is greater than a threshold (assuming it is 0.05). If so, it enters a sub-branch and continues to check the next condition; if not, it enters another sub-branch. This process continues until a leaf node is reached.

[0104] Assume that after a series of judgments, the decision tree determines that the RPA customer service text to be entered is an abnormal conversation. This judgment result is the abnormal conversation identification result for the entire text.

[0105] At the same time, since the abnormal conversation recognition result includes the abnormal conversation recognition result of each text segmentation word in the RPA customer service text to be entered, the computer system will also make a separate judgment on each text segmentation word. For example, for the segmentation word "user", the classifier may make another judgment based on the relevant features in the implicit representation set (through the association relationship with other vectors when the implicit representation set was previously constructed). Assuming that for the segmentation word "user", according to the judgment rule of the decision tree, its appearance in this session is related to the overall abnormal conversation situation, then it is recorded as abnormal correlation in the abnormal conversation recognition result (it may be indicated by marking or assigning a specific value). The computer system uses the classifier to process the implicit representation set to obtain the abnormal conversation recognition result of the RPA customer service text to be entered. This result provides a basis for subsequent data entry based on the abnormal conversation recognition result (such as step S400), which helps to improve the accuracy and reliability of RPA automated data entry.

[0106] As an implementation method, the method may further include:

[0107] Step S500: Divide the implicit representations in the implicit representation set into decomposed implicit representations of different scales;

[0108] Step S600: enhancing the implicit representations in the implicit representation set based on the decomposed implicit representations at different scales to obtain an enhanced implicit representation set.

[0109] When the computer system executes step S500, the purpose is to divide the implicit representations in the implicit representation set into decomposed implicit representations of different scales, which helps to analyze the information in the implicit representation from multiple scales and mine richer semantic features.

[0110] For example, suppose the RPA customer service text to be entered is "The customer reported that product M had a malfunction in a specific function, resulting in an interruption in the workflow." The implicit representation set obtained in the previous step includes the semantic representation vector of the RPA customer service text to be entered at the first level as [0.1, 0.2, -0.3, 0.4, -0.5], and the enhanced representation vector at the second level as [0.3, -0.1, 0.2, -0.3, 0.4] (this is a simplified example, and the dimension is set to 5). The semantic representation vector of the RPA customer service text to be entered at the first level and the enhanced representation vectors at x levels are scaled and decomposed respectively, and the semantic representation vector at the first level is operated first. The computer system extracts features based on the first convolution kernel and the second convolution kernel respectively. Assume that the first convolution kernel is (dimension is 3×3), the second convolution kernel is (The dimension is 5×5). For the first convolution kernel , the semantic representation vector is regarded as a two-dimensional feature map (here we simply convert the one-dimensional vector into two dimensions, for example, [0.1, 0.2, -0.3, 0.4, -0.5] is regarded as a 1×5 feature map). When performing the convolution operation, according to the calculation rules of convolution, for each position where the convolution kernel slides on the feature map, the sum of the products of the convolution kernel elements and the corresponding feature map elements is calculated. For example, at the first position, the calculated value is (here Represents the convolution kernel The element of the i-th row and j-th column) is calculated in this way to obtain the fine-grained extraction result at the first level. Then the fine-grained extraction result is pooled (for example, maximum pooling, taking the maximum value in the local area) to obtain the first fine-grained decomposition implicit representation. Similarly, for the second convolution kernel, the first coarse-grained extraction result is obtained through similar convolution and pooling operations, and then the first coarse-grained decomposition implicit representation is obtained through pooling. Then the same operation is performed on the enhanced representation vector at the e+1th level (here it is assumed that e=1, that is, the enhanced representation vector at the second level). Feature extraction is performed based on the first convolution kernel and the second convolution kernel respectively to obtain the fine-grained extraction result and the coarse-grained extraction result, and then pooling is performed respectively to obtain the e+1th fine-grained decomposition implicit representation and the e+1th coarse-grained decomposition implicit representation. For example, for the enhanced representation vector [0.3, -0.1, 0.2, -0.3, 0.4] at the second level, the corresponding fine-grained and coarse-grained decomposition implicit representations are obtained through operations similar to the above.

[0111] For the mth fine-grained decomposition implicit representation and the mth coarse-grained decomposition implicit representation (here we assume m = 1, i.e. the first fine-grained and coarse-grained decomposition implicit representations), they are first processed based on the feature extraction subcomponent. Assume that the feature extraction subcomponent is a small neural network, such as a multilayer perceptron (MLP) with an input layer, a hidden layer, and an output layer. For the fine-grained decomposition implicit representation (assumed to be [0.1, 0.2, -0.3], a simplified dimensional example), the number of neurons in the input layer is the same as the dimension of the fine-grained decomposition implicit representation, the hidden layer is set to 2 neurons, and the number of neurons in the output layer is set according to demand. Assume that the weight matrix from the input layer to the hidden layer is (dimension is 3×2), the bias vector is (The dimension is 2 dimensions), the weight matrix from the hidden layer to the output layer is (dimension is 2×2), the bias vector is (The dimension is 2 dimensions). The calculation process is as follows: First, calculate the output of the hidden layer ; (Here ReLU(x)=max(0,x)), and then calculate the output of the output layer as , the attention implicit representation of the m-th fine-grained decomposition implicit representation is obtained. Similarly, similar operations are performed on the coarse-grained decomposition implicit representation to obtain its attention implicit representation. Then, the obtained attention implicit representation is enhanced at the channel level based on the multi-head focusing subcomponent. Assume that two heads are used, and each head has a different weight matrix. For the attention implicit representation of the fine-grained decomposition implicit representation, two intermediate results are obtained by calculating with the weight matrix of each head (the calculation method is similar to the calculation in the previous multi-layer perceptron), and then these two results are fused at the channel level (for example, splicing or weighted summation, etc.) to obtain the m-th fine-grained transition feature. The m-th coarse-grained transition feature is obtained in the same way. Finally, based on the position focusing subcomponent, the m-th fine-grained transition feature and the m-th coarse-grained transition feature are enhanced at the position level. Assume that the position focusing subcomponent is strengthened through a position weight matrix W pos (The dimension is determined according to the dimension of the transition feature) The transition feature is weighted to obtain the mth enhanced fine-grained decomposition implicit representation and the mth enhanced coarse-grained decomposition implicit representation. When the computer system executes step S600, the implicit representation in the implicit representation set is enhanced by using the decomposition implicit representations of different scales obtained in step S500, thereby obtaining the enhanced implicit representation set.

[0112] For example, assume that three enhanced fine-grained decomposition implicit representations are obtained after step S500, namely [0.2, 0.3, -0.1], [0.1, -0.2, 0.3] and [-0.1, 0.2, 0.1]. The computer system merges these decomposition implicit representations based on the merge branch component. The merge branch component includes a local linear fitting component. Assume that the local linear fitting component works in the following manner: merge the s-th loading data and the x-s+1-th enhanced fine-grained decomposition implicit representation (here assuming x=3, s=1, then the x-s+1=3-th enhanced fine-grained decomposition implicit representation is [-0.1, 0.2, 0.1]).

[0113] First, based on the local linear fitting component, the s-th loading data (here it is assumed to be an initial value or the merged result of the previous round. If s=1, it can be a preset initial value, assumed to be [0, 0, 0]) is calibrated with the pooled data. The pooled data is obtained by pooling the x-s+1-th enhanced fine-grained decomposition implicit representation (for example, average pooling, the obtained value is [(-0.1+0.2+0.1) / 3, (-0.1+0.2+0.1) / 3, (-0.1+0.2+0.1) / 3]=[0.067,0.067, 0.067]). The calibration process may involve some calculations, such as calculating the difference between the two operations.

[0114] The displacement variable and the stretch variable are adjusted according to the calibration results. Assuming that the displacement variable is d and the stretch variable is s, the adjusted displacement variable d' and the adjusted stretch variable s' are calculated by the calibration results (the specific calculation method is based on the definition of the local linear fitting component and will not be described here).

[0115] Obtain the element-by-element product of the x-s+1th enhanced fine-grained decomposition implicit representation and the first expansion result to obtain a transition result. The first expansion result is obtained by expanding (such as upsampling) the adjusted stretched variable. Assuming that the adjusted stretched variable s'=2, the first expansion result is a vector obtained by multiplying each element by 2. Assuming that the x-s+1th enhanced fine-grained decomposition implicit representation is [-0.1, 0.2, 0.1], the transition result is [-0.1×2, 0.2×2,0.1×2]=[-0.2, 0.4, 0.2].

[0116] The transition result and the second expansion result are summed element by element to obtain the sth merged result. The second expanded result is obtained by expanding the adjusted displacement variable. Assuming that the adjusted displacement variable d'=1, the second expanded result is a vector obtained by adding 1 to each element, that is, [1, 1, 1]. Therefore, the sth merged result is [-0.2+1, 0.4+1, 0.2+1]=[0.8, 1.4, 1.2]. In this way, all the enhanced fine-grained decomposition implicit representations are gradually merged to obtain a set of fine-grained enhanced representation vectors. Similarly, the enhanced coarse-grained decomposition implicit representations are merged to obtain a set of coarse-grained enhanced representation vectors.

[0117] Assume that the feature extraction network is a simple linear layer neural network. For the set of fine-grained enhanced representation vectors (assumed to be [0.8, 1.4, 1.2], a simplified example) and the semantic representation vector [0.1, 0.2, -0.3] at the first level, the weight matrix of the linear layer is (dimension is 3×2), the bias vector is (The dimension is 2 dimensions). The calculation integration results are as follows: First, the set of fine-grained enhanced representation vectors and the semantic representation vectors are concatenated column by column into a new vector (assuming it is [0.8, 1.4, 1.2, 0.1, 0.2, -0.3]), and then this new vector is multiplied by the weight matrix and then added with the bias vector to obtain the first integration result. Similarly, for the set of coarse-grained enhanced representation vectors (assuming it is [0.5, 0.3, 0.1]) and the enhanced representation vector at the x-th level (assuming it is [0.3, -0.1, 0.2]), similar calculations are performed using another weight matrix (dimension is 3×2) and the bias vector (The dimension is 2) to obtain the second integration result. These first integration results and second integration results together constitute the enhanced implicit representation set.

[0118] As an implementation manner, step S500, dividing the implicit representations in the implicit representation set into decomposed implicit representations of different scales, may include:

[0119] Step S510: Decomposing the semantic representation vector of the RPA customer service text to be entered at the first level and the enhanced representation vector of the RPA customer service text to be entered at x levels respectively to obtain x+1 fine-grained decomposition implicit representations and x+1 coarse-grained decomposition implicit representations;

[0120] Step S520: Based on the dynamic focusing component, x+1 fine-grained decomposition implicit representations and x+1 coarse-grained decomposition implicit representations are enhanced respectively to obtain x+1 enhanced fine-grained decomposition implicit representations and x+1 enhanced coarse-grained decomposition implicit representations.

[0121] When the computer system executes step S510, it aims to decompose the vectors of the RPA customer service text to be entered at different levels through specific calculation operations, so as to obtain fine-grained and coarse-grained decomposition implicit representations. This helps to analyze the semantic information of the text from different levels of detail.

[0122] Assume that the RPA customer service text to be entered is "Customers complain that the operating interface of product N is too complicated and difficult to understand". In the previous step, the semantic representation vector of the text at the first level is [0.1, 0.2, -0.3, 0.4, -0.5] (to simplify the example, assume the dimension is 5), and the enhanced representation vector at the second level is [0.3, -0.1, 0.2, -0.3, 0.4] (here assume x=1).

[0123] The semantic representation vector at the first level is decomposed, and the semantic representation vector at the first level to be input into the RPA customer service text is extracted based on the first convolution kernel and the second convolution kernel, respectively, to obtain the fine-grained extraction result and the coarse-grained extraction result of the RPA customer service text at the first level. Assume that the first convolution kernel is (dimension is 2×2), the second convolution kernel is (The dimension is 3×3). For the first convolution kernel , the semantic representation vector is regarded as a two-dimensional feature map (here we simply convert the one-dimensional vector into two dimensions, for example, [0.1, 0.2, -0.3, 0.4, -0.5] is regarded as a 1×5 feature map). When performing the convolution operation, according to the calculation rules of convolution, for each position where the convolution kernel slides on the feature map, the sum of the products of the convolution kernel elements and the corresponding feature map elements is calculated. For example, at the first position (assuming the filling method is SAME filling), the calculated value is (here and They are The two elements in the first row are obtained by such calculation. Assume that the fine-grained extraction result obtained after the convolution operation is [0.05, 0.1, -0.15, 0.2] (the dimension is 4, simplified example). For the second convolution kernel , and perform convolution operations in the same way. Assume that the coarse-grained extraction result is [0.1, -0.2, 0.3] (the dimension is 3 dimensions, simplified example). Pool the fine-grained extraction results and coarse-grained extraction results of the RPA customer service text to be entered at the first level respectively to obtain the first fine-grained decomposition implicit representation and the first coarse-grained decomposition implicit representation. For the fine-grained extraction result [0.05, 0.1, -0.15, 0.2], assume that the maximum pooling operation is used (the pooling window size is 2×1). The first pooling takes the maximum value 0.1 between 0.05 and 0.1, and the second pooling takes the maximum value 0.2 between -0.15 and 0.2, and the first fine-grained decomposition implicit representation is [0.1, 0.2]. For the coarse-grained extraction result [0.1, -0.2, 0.3], assuming that the average pooling operation is used (the pooling window size is 3×1), (0.1 - 0.2+0.3) / 3 = 0.067 is calculated, and the first coarse-grained decomposition implicit representation is [0.067] (simplified to one-dimensional representation).

[0124] Next, the enhanced representation vector at the e+1th level is scaled (here e = 0 corresponds to the second level), and the enhanced representation vector at the e+1th level to be entered of the RPA customer service text is feature extracted based on the first convolution kernel and the second convolution kernel, respectively, to obtain the fine-grained extraction result and the coarse-grained extraction result of the RPA customer service text to be entered at the e+1th level. For the enhanced representation vector [0.3, -0.1, 0.2, -0.3, 0.4] at the second level, the first convolution kernel is used. (The dimension is 2×2) for convolution operation. Assume that the fine-grained extraction result is [0.1, 0.2, -0.1, 0.3] (the dimension is 4, simplified example). Use the second convolution kernel (dimension is 3×3) and the coarse-grained extraction result is assumed to be [0.2, -0.3, 0.4]. The fine-grained extraction result and the coarse-grained extraction result of the RPA customer service text to be entered at the e+1th level are pooled respectively to obtain the e+1th fine-grained decomposition implicit representation and the e+1th coarse-grained decomposition implicit representation. For the fine-grained extraction result [0.1, 0.2, -0.1,0.3], the maximum pooling operation (pooling window size is 2×1) is used to obtain the e+1th fine-grained decomposition implicit representation as [0.2, 0.3]. For the coarse-grained extraction result [0.2, -0.3, 0.4], the average pooling operation (pooling window size is 3×1) is used to obtain the e+1th coarse-grained decomposition implicit representation as [0.1] (simplified to a one-dimensional representation). In this way, the computer system obtains x+1 (here x=1, so 2) fine-grained decomposition implicit representations ([0.1, 0.2] and [0.2, 0.3]) and x+ 1 coarse-grained decomposition implicit representations ([0.067] and [0.1]).

[0125] When the computer system executes step S520, the dynamic focusing component is used to enhance the previously obtained fine-grained and coarse-grained decomposition implicit representations to further mine the semantic information therein and optimize the representation form.

[0126] Based on the feature extraction subcomponent, the mth fine-grained decomposition implicit representation and the mth coarse-grained decomposition implicit representation are processed respectively to obtain the attention implicit representation of the mth fine-grained decomposition implicit representation and the attention implicit representation of the mth coarse-grained decomposition implicit representation (here it is assumed that m = 1, that is, the first fine-grained and coarse-grained decomposition implicit representations are operated). Assume that the feature extraction subcomponent is a small neural network, such as a multilayer perceptron (MLP) comprising an input layer, a hidden layer, and an output layer. For the fine-grained decomposition implicit representation (assumed to be [0.1, 0.2]), the number of neurons in the input layer is the same as the dimension of the fine-grained decomposition implicit representation, that is, 2 neurons. The hidden layer is set to 1 neuron, and the number of neurons in the output layer is set to 1 neuron as required. Assume that the weight matrix from the input layer to the hidden layer is (dimension is 2×1), the bias vector is (The dimension is 1 dimension), the weight matrix from the hidden layer to the output layer is (dimension is 1×1), the bias vector is (The dimension is 1). Calculate the output of the hidden layer (Here ReLU(x)=max(0,x)). Assume ,but The output of the output layer is calculated as Assumptions , the output is , the attention implicit representation of the mth fine-grained decomposition implicit representation is obtained as [0.255]. For the coarse-grained decomposition implicit representation (assuming it is [0.067]), a similar MLP structure is also constructed. Assume that the weight matrix from the input layer to the hidden layer is (dimension is 1×1), the bias vector is (The dimension is 1 dimension), the weight matrix from the hidden layer to the output layer is (dimension is 1×1), the bias vector is (The dimension is 1 dimension). After similar calculation, the attention implicit representation of the mth coarse-grained decomposition implicit representation is obtained (assuming it is [0.3]).

[0127] Next, based on the multi-head focusing subcomponent, the attention implicit representation of the m-th fine-grained decomposition implicit representation and the attention implicit representation of the m-th coarse-grained decomposition implicit representation are enhanced at the channel level to obtain the m-th fine-grained transition feature and the m-th coarse-grained transition feature. Assume that two heads are used for multi-head focusing. For the attention implicit representation of the fine-grained decomposition implicit representation [0.255], the weight matrix of the first head is (dimension is 1×1), the weight matrix of the second head is (The dimension is 1×1). For the first head, the calculation result is [0.255]× , assuming =0.8, the result is [0.204]. For the second head, the result is [0.255]× , assuming =0.6, the result is [0.153]. The two results are fused (e.g., simply added) to obtain the mth fine-grained transition feature [0.357]. For the attention implicit representation [0.3] of the coarse-grained decomposition implicit representation, follow the same method, assuming that the first head weight matrix is = 0.9, the second head weight matrix is =0.7, and the mth coarse-grained transition feature is obtained through calculation (assumed to be [0.48]).

[0128] Next, based on the position focusing subcomponent, the mth fine-grained transition feature and the mth coarse-grained transition feature are enhanced at the position level to obtain the mth enhanced fine-grained decomposition implicit representation and the mth enhanced coarse-grained decomposition implicit representation. Assume that the position focusing subcomponent is represented by a position weight matrix (dimension is 1×1) weighted calculation of transition features. For the fine-grained transition feature [0.357], calculate [0.357]× Assumptions =1.2, the mth enhanced fine-grained decomposition implicitly expressed as [0.4284]. For the coarse-grained transition feature [0.48], assuming = 1.1, and the m-th enhanced coarse-grained decomposition is implicitly expressed as [0.528].

[0129] In this way, the computer system processes all x+1 fine-grained decomposition implicit representations and x+1 coarse-grained decomposition implicit representations respectively to obtain x+1 enhanced fine-grained decomposition implicit representations and x+1 enhanced coarse-grained decomposition implicit representations.

[0130] As an implementation mode, step S510 performs scale decomposition on the semantic representation vector of the RPA customer service text to be entered at the first level and the enhanced representation vector of the RPA customer service text to be entered at x levels, respectively, to obtain x+1 fine-grained decomposition implicit representations and x+1 coarse-grained decomposition implicit representations, which may include:

[0131] Step S511: extract features of the semantic representation vector of the RPA customer service text to be entered at the first level based on the first convolution kernel and the second convolution kernel respectively, and obtain a fine-grained extraction result and a coarse-grained extraction result of the RPA customer service text to be entered at the first level;

[0132] Step S512: pooling the fine-grained extraction results and the coarse-grained extraction results of the RPA customer service text to be input at the first level respectively to obtain the first fine-grained decomposition implicit representation and the first coarse-grained decomposition implicit representation;

[0133] Step S513: feature extraction is performed on the enhanced representation vector of the RPA customer service text to be entered at the e+1th level based on the first convolution kernel and the second convolution kernel respectively, to obtain the fine-grained extraction result and the coarse-grained extraction result of the RPA customer service text to be entered at the e+1th level, 1≤e≤x;

[0134] Step S514: Pool the fine-grained extraction results and the coarse-grained extraction results of the RPA customer service text to be entered at the e+1th level respectively to obtain the e+1th fine-grained decomposition implicit representation and the e+1th coarse-grained decomposition implicit representation.

[0135] In step S511, it is assumed that the semantic representation vector of the RPA customer service text to be entered at the first level is , whose dimension is n×m (for example, n=100, m=50, where n can represent the number of elements in the vector and m can represent the feature dimension of each element). The first convolution kernel The size of the second convolution kernel is a×b (for example, a=3, b=3). The size of is c×d (e.g. c=5, d=5).

[0136] For the first convolution kernel For fine-grained feature extraction, the computer system will follow the rules of convolution operation to exist Slide up to calculate. Take a simple one-dimensional convolution example (this is just for the convenience of understanding the concept, it is actually multi-dimensional), assuming =[1,2,3,4,5], =[1,1], then the first element of the fine-grained feature extraction result is calculated as ((1×1+ 2×1)=3), and so on to calculate the entire vector to obtain the fine-grained extraction result .

[0137] For the second convolution kernel The coarse-grained feature extraction also follows the convolution operation rules, because The larger the size, the more elements will be covered for calculation, resulting in a coarse-grained extraction result. The use of convolution kernels of different sizes can capture features of different scales, just like using filters of different sizes to filter information. Small filters (the first convolution kernel) can capture more detailed features, and large filters (the second convolution kernel) can capture more macroscopic features.

[0138] The computer system obtains the fine-grained extraction result in step S511 And coarse-grained extraction results After that, pooling operation is required. Assume The dimension is p×q (e.g. p=50, q=30), The dimension is r×s (e.g. r=30, s=20).

[0139] Taking the maximum pooling as an example, for fine-grained extraction results , if the pooling window size is 2×2, the computer system will In the matrix, the maximum value in each window is selected as the pooling result in units of 2×2 windows. For example, the value in the window is , then the result after pooling is 4. In this way, the entire Pooling is performed to obtain the first fine-grained decomposition implicit representation .

[0140] Similarly, for the coarse-grained extraction results , according to the same pooling method (for example, also maximum pooling, the pooling window size can be different, assuming it is 3×3, the operation is performed to obtain the first coarse-grained decomposition implicit representation The pooling operation can further reduce the amount of data while retaining the main feature information, just like selecting the most significant part from a large image to represent the image. In step S513, for each e+1 level The enhanced representation vector , assuming its dimension is u×v (e.g. u=80, v=40). The computer system also uses the first convolution kernel And the second convolution kernel Perform feature extraction.

[0141] Taking e=1 as an example, for the enhanced representation vector According to the convolution operation rules in step S511, the first convolution kernel is used Perform fine-grained feature extraction. Assume = [2,3,4,5,6,7,8,9], =[1, -1], then the first element of the fine-grained feature extraction result is calculated as , and so on to calculate the fine-grained extraction results .

[0142] Use the second convolution kernel When performing coarse-grained feature extraction, due to Larger coverage, coarse-grained extraction results For other e-value levels, the computer system also operates in the same way to obtain fine-grained and coarse-grained extraction results at each level.

[0143] In step S514, the computer system obtains the fine-grained extraction result in step S513. And coarse-grained extraction results After that, the pooling operation is performed. Assume The dimension is w×z (for example, w=40, z=25, The dimensions are x×y (e.g. x=25, y=15).

[0144] Taking average pooling as an example, for fine-grained extraction results , if the pooling window size is 2×2, the computer system will In the matrix, the average value of the elements in each window is calculated as the result of pooling, with 2×2 windows as the unit. For example, the value in the window is , then the result after pooling is (1 + 2+3 + 4) / 4=2.5. Pooling is performed to obtain the e+1th fine-grained decomposition implicit representation .

[0145] Similarly, for the coarse-grained extraction results , according to the same pooling method (for example, average pooling, the pooling window size can be different, assuming it is 3×3, and the operation is performed to obtain the e+1th coarse-grained decomposition implicit representation . Through such a series of operations, the computer system can obtain fine-grained and coarse-grained decomposed implicit representations for the vectors at different levels of the RPA customer service text to be entered. These representations will be used in the subsequent steps to perform operations such as strengthening the implicit representations in the implicit representation set, which will help improve the accuracy and effectiveness of the entire RPA automated data entry method. In the entire RPA automated data entry scenario, it is assumed that the RPA customer service text is about the customer's consultation content on a certain product, such as "The product I bought has a fault and the screen flickers. What should I do?" When performing the above operations, the processing of these vectors can better understand the different levels and features of the text semantics, so as to accurately identify abnormal conversations and perform correct data entry. For example, if it is found that the screen flickering-related features have special performance in the fine-grained decomposed implicit representation of a certain level in the abnormal conversation identification, more accurate processing can be performed based on this information, such as classifying the problem as an abnormal conversation type related to hardware failure, thereby providing an accurate basis for subsequent data entry.

[0146] As an implementation, the dynamic focusing component includes a feature extraction subcomponent, a position focusing subcomponent, and a multi-head focusing subcomponent. Based on this, step S520, based on the dynamic focusing component, performs enhancement processing on x+1 fine-grained decomposition implicit representations and x+1 coarse-grained decomposition implicit representations respectively to obtain x+1 enhanced fine-grained decomposition implicit representations and x+1 enhanced coarse-grained decomposition implicit representations, which may include:

[0147] Step S521: Based on the feature extraction subcomponent, the m-th fine-grained decomposition implicit representation and the m-th coarse-grained decomposition implicit representation are processed to obtain the attention implicit representation of the m-th fine-grained decomposition implicit representation and the attention implicit representation of the m-th coarse-grained decomposition implicit representation, m≤(x+1);

[0148] Step S522: Based on the multi-head focusing subcomponent, the attention implicit representation of the m-th fine-grained decomposition implicit representation and the attention implicit representation of the m-th coarse-grained decomposition implicit representation are enhanced at the channel level to obtain the m-th fine-grained transition feature and the m-th coarse-grained transition feature;

[0149] Step S523: Based on the position focusing subcomponent, the mth fine-grained transition feature and the mth coarse-grained transition feature are enhanced at the position level to obtain the mth enhanced fine-grained decomposition implicit representation and the mth enhanced coarse-grained decomposition implicit representation.

[0150] In the embodiment of the present application, step S520 uses a dynamic focusing component to enhance the previously obtained fine-grained decomposition implicit representation and coarse-grained decomposition implicit representation. The following will explain in detail the sub-steps included in this step.

[0151] In step S521, assuming that in the RPA automated data entry scenario, for the mth fine-grained decomposition implicit representation , whose dimension is a×b, can be expressed as a feature vector matrix, in which each element contains the fine-grained feature information about the RPA customer service text to be entered at a specific level. Coarse-grained decomposition implicitly represents , whose dimension is c×d.

[0152] Taking a feature extraction subcomponent based on a convolutional neural network (CNN) as an example, it may contain multiple convolutional layers and activation functions. , the computer system first inputs it into the feature extraction subcomponent. Assume that the convolution kernel size of the convolution layer is 3×3, the stride is 1, and the padding is SAME. When passing through the first convolution layer, the calculation is as follows: For each element in the output matrix (i and j represent the row and column indices of the matrix respectively), ,in is the weight of the convolution kernel, is the input matrix The corresponding element in , b is the bias term. After being processed by multiple convolutional layers and activation functions (such as ReLU function, i.e. (f(x)=max(0,x))), the attention implicit representation of the mth fine-grained decomposition implicit representation is obtained .

[0153] Similarly, for the coarse-grained decomposition implicit representation , the computer system follows the same process and processes through the feature extraction subcomponent to obtain the attention implicit representation of the mth coarse-grained decomposition implicit representation This process is similar to deep mining of the original fine-grained and coarse-grained features to obtain more representative and interesting feature representations, just like screening out those parts that are more critical to subsequent processing (such as abnormal conversation identification, etc.) from a pile of feature information about customer service texts.

[0154] In step S522, assume that the attention implicit representation of the m-th fine-grained decomposition implicit representation is The number of channels is n (e.g. n=16), and the attention implicit representation of the mth coarse-grained decomposition implicit representation is The number of channels is m (for example, m=8).

[0155] Taking the Multi-Head Attention mechanism as an example, the multi-head focus subcomponent may contain h heads (for example, h=4). , first split it into h sub-representations, each with a dimension of (here yes For each sub-representation, calculate its attention weight. Assume that Q (query matrix), K (key matrix) and V (value matrix) are obtained by linearly transforming each sub-representation, and their dimensions are ;For example =10, =20. Attention weight , and then by The output of each head is calculated. Finally, the outputs of h heads are concatenated to obtain the mth fine-grained transition feature after channel-level enhancement. .

[0156] Attention implicit representation for the mth coarse-grained decomposition implicit representation , the computer system follows the same multi-head attention mechanism calculation process to obtain the mth coarse-grained transition feature This enhanced processing at the channel level enables the model to pay more attention to the relationship between different channels and further explore the correlation information between features. Just like in the feature representation of RPA customer service text, it can gain a deeper understanding of the interaction between different feature dimensions (channels) in order to better process the text.

[0157] In step S523, the computer system uses the position focusing subcomponent to perform processing in this step. Assume that the mth fine-grained transition feature The dimension is e×f (e.g. e=40, f=25), and the mth coarse-grained transition feature The dimension is g×h (e.g. g=20, h=15).

[0158] Take a position focusing subcomponent based on relative position encoding as an example. For the mth fine-grained transition feature , the computer system will add position codes according to the position information of the elements in the matrix. Assume that the dimension of the relative position code matrix P is the same as Similarly, the position encoding can be calculated as ,in is the result of adding position encoding. Then a linear transformation layer (such as a fully connected layer) is used to transform After processing, the mth enhanced fine-grained decomposition implicit representation is obtained .

[0159] Similarly, for the mth coarse-grained transition feature The computer system follows the same process, first adding the relative position encoding, and then processing it through the linear transformation layer to obtain the mth enhanced coarse-grained decomposition implicit representation This enhanced processing at the position level takes into account the position information of feature elements. In the RPA customer service text processing scenario, it can better capture the sequential relationship of features in the text. For example, when identifying abnormal conversations in customer service texts, it can more accurately determine whether the position of a feature in the text conforms to normal semantic logic, thereby improving the accuracy and reliability of the entire RPA automated data entry method.

[0160] Through the series of operations in the above step S520, the computer system can effectively enhance the fine-grained and coarse-grained decomposition implicit representations at different levels, and these enhanced representations will continue to be used for operations on implicit representation sets in subsequent steps, further improving the performance of the entire RPA automated data entry process. For example, when processing the RPA customer service text to be entered, "The product suddenly heated up during use, and I am worried whether it will explode. This has never happened before", the feature representations after these enhanced processing can more accurately identify the semantic relationship and importance of key information such as "suddenly heated up" and "worried about explosion" in the text, which helps to accurately determine whether this is an abnormal session and perform correct data entry.

[0161] As an implementation manner, the decomposed implicit representations at different scales include x+1 enhanced fine-grained decomposed implicit representations and x+1 enhanced coarse-grained decomposed implicit representations; based on this, step S600, based on the decomposed implicit representations at different scales, performs enhancement processing on the implicit representations in the implicit representation set to obtain the enhanced implicit representation set, which may include:

[0162] Step S610: merging x+1 enhanced fine-grained decomposition implicit representations to obtain a set of fine-grained enhanced representation vectors, and merging x+1 enhanced coarse-grained decomposition implicit representations to obtain a set of coarse-grained enhanced representation vectors;

[0163] Step S620: Integrate the fine-grained enhanced representation vector set and the semantic representation vector of the RPA customer service text to be entered at the first level based on the feature extraction network to obtain a first integration result;

[0164] Step S630: Based on the feature extraction network, the coarse-grained enhanced representation vector set and the enhanced representation vector at the xth level of the RPA customer service text to be entered are integrated to obtain a second integration result.

[0165] In the present RPA automated data entry method combined with artificial intelligence, step S600 aims to enhance the implicit representations in the implicit representation set based on the decomposed implicit representations of different scales, so as to obtain the enhanced implicit representation set. The sub-steps included in this step will be explained in detail below.

[0166] Step S610 performs a merging operation on multiple enhanced decomposition implicit representations. Assuming that in the RPA automated data entry scenario, x=3, there are (x+1=4) enhanced fine-grained decomposition implicit representations, which are respectively recorded as .

[0167] by For example, assuming its dimension is (For example = 30, = 20), is a matrix containing specific fine-grained feature information about the RPA customer service text to be entered. For the merging of these enhanced fine-grained decomposition implicit representations, the computer system can use element-by-element addition (this is only one possible merging method). For example, for the first element position in the matrix (i=0, j=0), the fine-grained enhanced representation vector set The value of the element (at this position) is calculated as In this way, the entire matrix is ​​calculated to obtain a set of fine-grained enhanced representation vectors , which is a feature map that integrates fine-grained feature information at different levels.

[0168] Similarly, for the x+1 enhanced coarse-grained decomposition implicit representations, assume that they are , whose dimensions may be (For example =20, = 15). The computer system merges them in a similar way (e.g., element-by-element addition) to obtain a set of coarse-grained enhanced representation vectors. This merging operation can integrate feature information at different levels, just like when analyzing RPA customer service text, the coarse-grained and fine-grained features scattered at various levels are integrated to facilitate more comprehensive text processing later. In step S620, the feature extraction network is used to integrate information. Assume that the semantic representation vector of the RPA customer service text to be entered at the first level is , the dimension is (For example ), the set of fine-grained enhanced representation vectors obtained above The dimension is (For example ).

[0169] Take a simple fully connected neural network (FCNN) as an example of a feature extraction network. First, the computer system will and Perform the concatenation operation to obtain a new vector Z. Assuming that the concatenation method is to connect in dimension, the dimension of Z is .

[0170] Then Z is input into the fully connected neural network. Each layer of the fully connected neural network contains multiple neurons. Assume that the first layer has neurons (e.g. = 100). For each element in the input Z , and the connection weight with the first layer of neurons is ), the bias term is , then the output of the first layer of neurons is , after an activation function (such as the ReLU function After processing, the intermediate result is obtained. After calculations of multiple layers of fully connected layers, the first integrated result is finally obtained .

[0171] In the RPA automated data entry scenario, for example, if the customer service text to be entered is about a product failure description, the semantic representation vector at the first level may contain basic semantic information about the text, while the set of fine-grained enhanced representation vectors contains fine-grained features extracted and enhanced from different levels. This integration enables a more comprehensive understanding of the semantics and features of the text, which helps to accurately identify abnormal conversations and other operations later.

[0172] Step S630 is similar to step S620, and also uses the feature extraction network to perform integration operations. Assume that the enhanced representation vector of the RPA customer service text to be entered at the xth level (here x = 3) is , whose dimensions are (For example ), the coarse-grained enhanced representation vector set The dimension is (For example ).

[0173] Also using the fully connected neural network as the feature extraction network, the computer system first and Perform the concatenation operation to obtain a new vector W, whose dimension is Then W is input into the fully connected neural network, and the calculation method is similar to that in step S620, that is, for each element in the input W , and the connection weight with the first layer of neurons is ),here can be combined with the step S620 Same or different, e.g. , the bias term is , then the output of the first layer of neurons is After being processed by the activation function and calculated by multiple layers of fully connected layers, the second integration result is finally obtained .

[0174] In RPA customer service text processing, for example, when the customer service text describes a complex problem of a product, the enhanced representation vector at the xth level may contain higher-level semantic and feature information, and the coarse-grained enhanced representation vector set contains coarse-grained features integrated from different levels. Through this integration, the deep semantic relationship in the text can be further mined, and the ability to understand the text can be improved, so that various situations can be handled more accurately during the data entry process, such as abnormal session recognition and accurate entry of relevant data. Through the operation of the above step S600, the computer system can effectively use the decomposed implicit representations of different scales to enhance the implicit representations in the implicit representation set, and obtain the enhanced implicit representation set, which helps to improve the accuracy and effectiveness of the entire RPA automated data entry method.

[0175] As an implementation manner, step S610, merging x+1 enhanced fine-grained decomposition implicit representations to obtain a set of fine-grained enhanced representation vectors, may include:

[0176] Step S611: Based on the merge branch component, merge the sth loaded data and the x-s+1th enhanced fine-grained decomposition implicit representation to obtain the sth merged result; wherein, 1≤s≤x; if s=1, the sth loaded data is the x+1th enhanced fine-grained decomposition implicit representation; if 1<s<x+1, the sth loaded data is the s-1th merged result.

[0177] In the present RPA automated data entry method combined with artificial intelligence, step S610 aims to merge the \(x + 1\) enhanced fine-grained decomposition implicit representations to obtain a set of fine-grained enhanced representation vectors. Among them, step S611 describes in detail the specific method of merging.

[0178] When the computer system executes this step, it involves merging data based on the merge branch component. Assuming that in the RPA automated data entry scenario, when the computer system executes this step, it involves merging data based on the merge branch component. Assuming that in the RPA automated data entry scenario, there is The enhanced fine-grained decomposition implicit representations are denoted as .

[0179] by For example, assuming its dimension is (For example ), these dimensions represent the structure of the feature information contained in the implicit representation. When the sth loading data is The enhanced fine-grained decomposition implicit representation is .

[0180] Assume that the merged branch component includes a local linear fitting component. The computer system performs and Merge to obtain the \(1\)th merged result. Specifically, firstly, based on the local linear fitting component, Calibrate with the pooled data to obtain the calibration result. Perform a pooling operation (e.g., maximum pooling, with a pooling window size of Get the pooled data P. The calibration process involves complex calculations. Here, it is assumed that some predefined function \(f\) is used for calibration. The calibration result .

[0181] Then, the displacement variable is calculated based on the calibration results. and stretch variables Make adjustments to obtain the adjusted displacement variable and the adjusted stretch variable Assuming that the adjustment functions are g and h, then Then get The transition result is obtained by element-wise multiplication of the first expansion result. The first expansion result is obtained by adjusting the stretched variable Perform expansion (such as upsampling operation) to obtain, assuming that the expansion function is u, the first expansion result is , then the transition result .

[0182] Finally, the transition result T and the second expansion result are summed element by element to obtain the first combined result. The second expansion result is the adjusted displacement variable Expand and obtain, assuming that the expansion function is v, the second expansion result is , then the first merge result .

[0183] when (For example, when s=2), the sth loading data is the s-1th merged result, that is, \(M_1\). The computer system follows the same process based on the local linear fitting component. and Merge to obtain the second merge result. And so on, until all the merge operations are completed, finally a set of fine-grained enhanced representation vectors is obtained. This merging method can effectively integrate different fine-grained decomposition implicit representations. In the RPA automated data entry scenario, it helps to more comprehensively capture the feature information of customer service texts, thereby improving the accuracy and efficiency of data entry.

[0184] As an implementation mode, the merging branch component includes a local linear fitting component; based on this, step S611, based on the merging branch component, merges the s-th loaded data and the x-s+1-th enhanced fine-grained decomposition implicit representation to obtain the s-th merged result, which may include:

[0185] Step S6111: calibrating the s-th loaded data with the pooled data based on the local linear fitting component to obtain a calibration result, wherein the pooled data is obtained by pooling the x-s+1-th enhanced fine-grained decomposition implicit representation;

[0186] Step S6112: adjusting the displacement variable and the stretch variable according to the calibration result to obtain an adjusted displacement variable and an adjusted stretch variable;

[0187] Step S6113: obtaining the element-by-element product of the x-s+1th enhanced fine-grained decomposition implicit representation and the first expansion result to obtain a transition result, wherein the first expansion result is obtained by expanding the adjusted stretching variable;

[0188] Step S6114: sum the transition result and the second expansion result element by element to obtain the sth combined result, where the second expansion result is obtained by expanding the adjusted displacement variable.

[0189] In step S6111, the local linear fitting component plays a key role. Assume that , when s=3, the sth loaded data is the previously calculated data (here it is assumed to be a vector data in matrix form), recorded as , whose dimensions are (For example ), this data contains some feature information about the text to be entered into the RPA customer service accumulated in the previous processing steps.

[0190] For The third enhanced fine-grained decomposition implicit representation is denoted as , assuming its dimension is (For example ). The computer system first Perform a pooling operation to obtain pooled data. Taking the maximum pooling as an example, if the pooling window size is 2×2, for For each 2×2 submatrix in , select the maximum value as the pooling result. For example, the submatrix The value after pooling is 4. After the pooling operation, the pooled data P is obtained, and its dimension becomes (Here we assume that p and q are both even numbers for ease of understanding. In actual situations, they can be handled flexibly according to the specific pooling strategy and data dimension.) For example, it becomes .

[0191] The computer system then uses a local linear fitting component to and The local linear fitting component may be based on a predefined linear function model, for example, assuming that the model is In the form of (here x represents the element in P, y represents the result element after calibration, and a and b are predefined parameters). (i and j represent row and column indices respectively), the calibration result The calculation method is (Here we assume The elements at the same position as P have a certain correlation relationship for calibration). Through such calculation, the computer system obtains the calibration result C, which integrates the characteristic information of the s-th loading data and the pooling data, preparing for the subsequent variable adjustment.

[0192] In step S6112, after obtaining the calibration result C, the computer system adjusts the displacement variable and the stretch variable according to it. Assume that the displacement variable is , the initial value is (For example ), the stretching variable is , the initial value is (For example ).

[0193] During this process, the computer system may adjust the two variables using a mapping function based on the values ​​of the calibration result elements. For example, for the displacement variable The regulation function can be defined as ,in is a predefined weight coefficient (e.g. The value of is determined based on the importance of the data or other prior knowledge). Assume is a uniformly distributed weight matrix, whose element values ​​are all 0.01 (just for example convenience). The adjusted displacement variable can be obtained by calculation .

[0194] For the stretch variable The regulation function can be defined as ,in is another predefined weight coefficient (e.g. The value of is determined according to the characteristics of the data). Assume is a randomly generated weight matrix with element values ​​between [- 0.1, 0.1]. The adjusted stretch variable is obtained by calculation This adjustment method enables the displacement variables and stretch variables to be adaptively adjusted according to the characteristic information in the calibration results, so as to better process the data later.

[0195] In step S6113, firstly adjust the adjusted stretching variable Expand (upsample) to obtain the first expansion result. Assuming a simple linear upsampling method is adopted, for (assuming a scalar value), if we want to extend it to the implicit representation of the fine-grained decomposition after the x-s+1th enhancement The same dimensions (i.e. ), then the first expansion result Each element in It can be calculated as (This is a simple uniform expansion method, and in actual situations it can be processed according to a more complex upsampling algorithm).

[0196] The computer system then calculates the The enhanced fine-grained decomposition implicit representation With the first expanded result The element-wise product of . Each element in and The corresponding elements in , transition result Elements in Calculated as By this element-by-element multiplication, the transition result T integrates The information expanded with the stretching variable reflects the characteristic changes.

[0197] In step S6114, after obtaining the transition result T, the computer system needs to sum it with the second expansion result element by element to obtain the sth combined result. Perform expansion to obtain a second expansion result. Assuming that a similar expansion method as in step S6113 is adopted, Expand to The same dimension gets the second expansion result , whose elements Calculated as (Again a simple way of evenly expanding).

[0198] Then, for the transition result Each element in and the second extended result The corresponding elements in , the sth merge result Elements in Calculated as By summing up the elements one by one, the s-th merged result The transition result T and the expanded information of displacement variables are integrated to complete the The merging operation of the implicit representation of the enhanced fine-grained decomposition.

[0199] In the entire RPA automated data entry process, for example, for the customer service text to be entered, "Intermittent failures occur when the product is used, and the software frequently reports errors", the operations in the above steps help to better process the feature information in the text. The fine-grained decomposition implicit representations after reinforcement at different levels contain features about different aspects of the text. Through such merging operations, these features can be more comprehensively integrated to provide richer and more accurate feature representations for subsequent abnormal session recognition, data entry and other operations, thereby improving the performance and accuracy of the entire RPA automated data entry method.

[0200] The present application embodiment provides a computer system, such as Figure 2 As shown, the computer system 100 includes: a processor 101 and a memory 103. The processor 101 and the memory 103 are connected, such as through a bus 102. Optionally, the computer system 100 may also include a transceiver 104. It should be noted that in actual applications, the transceiver 104 is not limited to one, and the structure of the computer system 100 does not constitute a limitation on the embodiments of the present application.

[0201] An embodiment of the present application provides a computer system. The computer system in the embodiment of the present application includes: one or more processors; a memory; and one or more computer programs, wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors. When the one or more programs are executed by the processor, the above method of the embodiment of the present application is implemented.

Claims

1. A RPA automated data entry method combined with artificial intelligence, characterized in that: The method comprises: Obtaining a text semantic implicit representation and a word frequency implicit representation of the RPA customer service text to be entered; the text semantic implicit representation includes a semantic representation vector of the RPA customer service text to be entered at a first level and a semantic representation vector of the RPA customer service text to be entered at a second level; According to the text semantic implicit representation and the word frequency implicit representation, an implicit representation set of the RPA customer service text to be entered is obtained; the implicit representation set includes the semantic representation vector of the RPA customer service text to be entered at the first level and the enhanced representation vectors of the RPA customer service text to be entered at x levels, and the enhanced representation vector of the RPA customer service text to be entered at any level is obtained based on the semantic representation vector of the RPA customer service text to be entered at the second level and the word frequency implicit representation, and x≥1; According to the implicit representation set, abnormal session identification is performed on the RPA customer service text to be input, and an abnormal session identification result of the RPA customer service text to be input is obtained; Entering the RPA customer service text to be entered based on the abnormal session identification result; Wherein, obtaining the implicit representation set of the text to be input into the RPA customer service text according to the text semantic implicit representation and the word frequency implicit representation includes: Integrate the semantic representation vector of the RPA customer service text to be input at the second level and the word frequency implicit representation to obtain a reinforced word frequency implicit representation; Integrate the semantic representation vector of the RPA customer service text to be entered at the second level with the enhanced word frequency implicit representation to obtain the enhanced representation vector of the RPA customer service text to be entered at the second level; Obtaining an implicit representation set of the RPA customer service text to be entered according to the semantic representation vector of the RPA customer service text to be entered at the first level and the enhanced representation vector of the RPA customer service text to be entered at the second level; The method further comprises: Repeatedly embed and map the enhanced representation vector of the RPA customer service text to be entered at the second level based on y embedding mapping components to obtain the enhanced representation vector of the RPA customer service text to be entered at y levels, where y≥1; The step of obtaining an implicit representation set of the RPA customer service text to be entered based on the semantic representation vector of the RPA customer service text to be entered at the first level and the enhanced representation vector of the RPA customer service text to be entered at the second level includes: The semantic representation vector of the RPA customer service text to be entered at the first level, the enhanced representation vector of the RPA customer service text to be entered at the second level, and the enhanced representation vectors of the RPA customer service text to be entered at y levels are aggregated to obtain an implicit representation set of the RPA customer service text to be entered; The step of integrating the semantic representation vector of the RPA customer service text to be input at the second level and the word frequency implicit representation to obtain the enhanced word frequency implicit representation includes: Based on the position focus strategy, the semantic representation vector of the RPA customer service text to be input at the second level and the word frequency implicit representation are processed to obtain the first attention implicit representation; Based on the multi-head focusing strategy, the semantic representation vector of the RPA customer service text to be input at the second level and the word frequency implicit representation are processed to obtain the second attention implicit representation; Processing a fusion result of the first attention implicit representation and the second attention implicit representation based on a feature extraction component to obtain a reinforced word frequency implicit representation; The step of integrating the semantic representation vector of the RPA customer service text to be entered at the second level with the enhanced word frequency implicit representation to obtain the enhanced representation vector of the RPA customer service text to be entered at the second level includes: Performing feature extraction on the enhanced word frequency implicit representation based on a target feature extraction network to obtain a feature extraction result of the enhanced word frequency implicit representation; The feature extraction result implicitly represented by the enhanced word frequency is summed element by element with the semantic representation vector of the RPA customer service text to be entered at the second level to obtain the enhanced representation vector of the RPA customer service text to be entered at the second level.

2. The method according to claim 1, characterized in that The method further comprises: Dividing the implicit representations in the implicit representation set into decomposed implicit representations of different scales; Based on the decomposed implicit representations of different scales, the implicit representations in the implicit representation set are enhanced to obtain an enhanced implicit representation set.

3. The method according to claim 2, characterized in that The step of dividing the implicit representations in the implicit representation set into decomposed implicit representations of different scales comprises: The semantic representation vector of the RPA customer service text to be entered at the first level and the enhanced representation vector of the RPA customer service text to be entered at x levels are respectively decomposed to obtain x+1 fine-grained decomposition implicit representations and x+1 coarse-grained decomposition implicit representations; Based on the dynamic focusing component, the x+1 fine-grained decomposition implicit representations and the x+1 coarse-grained decomposition implicit representations are enhanced respectively to obtain x+1 enhanced fine-grained decomposition implicit representations and x+1 enhanced coarse-grained decomposition implicit representations.

4. The method according to claim 3, characterized in that The semantic representation vector of the RPA customer service text to be entered at the first level and the enhanced representation vector of the RPA customer service text to be entered at x levels are scaled separately to obtain x+1 fine-grained decomposition implicit representations and x+1 coarse-grained decomposition implicit representations, including: Based on the first convolution kernel and the second convolution kernel, feature extraction is performed on the semantic representation vector of the RPA customer service text to be entered at the first level, respectively, to obtain a fine-grained extraction result and a coarse-grained extraction result of the RPA customer service text to be entered at the first level; The fine-grained extraction results and the coarse-grained extraction results of the RPA customer service text to be input at the first level are pooled respectively to obtain the first fine-grained decomposition implicit representation and the first coarse-grained decomposition implicit representation; Based on the first convolution kernel and the second convolution kernel, feature extraction is performed on the enhanced representation vector of the RPA customer service text to be entered at the e+1th level, respectively, to obtain a fine-grained extraction result and a coarse-grained extraction result of the RPA customer service text to be entered at the e+1th level, 1≤e≤x; The fine-grained extraction results and the coarse-grained extraction results of the RPA customer service text to be input at the e+1th level are respectively pooled to obtain the e+1th fine-grained decomposition implicit representation and the e+1th coarse-grained decomposition implicit representation; The dynamic focusing component includes a feature extraction subcomponent, a position focusing subcomponent and a multi-head focusing subcomponent; the x+1 fine-grained decomposition implicit representations and the x+1 coarse-grained decomposition implicit representations are enhanced based on the dynamic focusing component to obtain x+1 enhanced fine-grained decomposition implicit representations and x+1 enhanced coarse-grained decomposition implicit representations, including: Based on the feature extraction subcomponent, the m-th fine-grained decomposition implicit representation and the m-th coarse-grained decomposition implicit representation are processed to obtain an attention implicit representation of the m-th fine-grained decomposition implicit representation and an attention implicit representation of the m-th coarse-grained decomposition implicit representation, m≤(x+1); Based on the multi-head focusing subcomponent, the attention implicit representation of the m-th fine-grained decomposition implicit representation and the attention implicit representation of the m-th coarse-grained decomposition implicit representation are enhanced at the channel level to obtain the m-th fine-grained transition feature and the m-th coarse-grained transition feature; Based on the position focusing subcomponent, the mth fine-grained transition feature and the mth coarse-grained transition feature are enhanced at the position level to obtain the mth enhanced fine-grained decomposition implicit representation and the mth enhanced coarse-grained decomposition implicit representation.

5. The method according to claim 2, characterized in that The decomposed implicit representations of different scales include x+1 enhanced fine-grained decomposed implicit representations and x+1 enhanced coarse-grained decomposed implicit representations; the step of enhancing the implicit representations in the implicit representation set based on the decomposed implicit representations of different scales to obtain an enhanced implicit representation set includes: Merging the x+1 enhanced fine-grained decomposition implicit representations to obtain a set of fine-grained enhanced representation vectors, and merging the x+1 enhanced coarse-grained decomposition implicit representations to obtain a set of coarse-grained enhanced representation vectors; Integrate the fine-grained enhanced representation vector set and the semantic representation vector of the RPA customer service text to be entered at the first level based on the feature extraction network to obtain a first integration result; Integrate the coarse-grained enhanced representation vector set and the enhanced representation vector of the RPA customer service text to be entered at the xth level based on the feature extraction network to obtain a second integration result; The x+1 enhanced fine-grained decomposition implicit representations are merged to obtain a set of fine-grained enhanced representation vectors, including: Based on the merging branch component, the sth loaded data and the x-s+1th enhanced fine-grained decomposition implicit representation are merged to obtain the sth merged result; wherein, 1≤s≤x; if s=1, the sth loaded data is the x+1th enhanced fine-grained decomposition implicit representation; if 1<s<x+1, the sth loaded data is the s-1th merged result.

6. The method according to claim 5, characterized in that The merging branch component includes a local linear fitting component; the merging of the s-th loading data and the x-s+1-th enhanced fine-grained decomposition implicit representation based on the merging branch component to obtain the s-th merging result includes: Calibrate the s-th loading data with the pooled data based on the local linear fitting component to obtain a calibration result, wherein the pooled data is obtained by pooling the x-s+1-th enhanced fine-grained decomposition implicit representation; Adjusting the displacement variable and the stretch variable according to the calibration result to obtain an adjusted displacement variable and an adjusted stretch variable; Obtaining an element-by-element product of the x-s+1th enhanced fine-grained decomposition implicit representation and a first expansion result to obtain a transition result, wherein the first expansion result is obtained by expanding the adjusted stretching variable; The transition result and the second expansion result are summed element by element to obtain the sth combined result, wherein the second expansion result is obtained by expanding the adjusted displacement variable.

7. The method according to claim 1, characterized in that The performing abnormal session identification on the RPA customer service text to be input according to the implicit representation set to obtain the abnormal session identification result of the RPA customer service text to be input includes: Classifying the implicit representation set based on the classifier to obtain an abnormal conversation recognition result of the RPA customer service text to be entered; the abnormal conversation recognition result includes an abnormal conversation recognition result of each text segmentation in the RPA customer service text to be entered; The obtaining of the text semantic implicit representation and word frequency implicit representation of the text to be entered into the RPA customer service text includes: Based on the first embedding mapping component, embedding mapping is performed on the RPA customer service text to be input, so as to obtain a semantic representation vector of the RPA customer service text to be input at the first level; Based on the second embedding mapping component, embedding and mapping the semantic representation vector of the RPA customer service text to be entered at the first level to obtain the semantic representation vector of the RPA customer service text to be entered at the second level; The semantic representation vector of the RPA customer service text to be entered at the first level and the semantic representation vector of the RPA customer service text to be entered at the second level are used as the text semantic implicit representation of the RPA customer service text to be entered; Performing word frequency statistics on the RPA customer service text to be entered through a word frequency channel to obtain a word frequency statistics result of the RPA customer service text to be entered; The word frequency statistics result of the RPA customer service text to be entered is quantified based on the word frequency table to obtain an implicit representation of the word frequency of the RPA customer service text to be entered.

8. A computer system, characterized in that: include: one or more processors; Memory; one or more computer programs; The one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, the method according to any one of claims 1 to 7 is implemented.

Images