A customer service result prediction model establishment method, device and equipment

Abstract

This invention provides a method, apparatus, and device for establishing a customer service outcome prediction model, relating to the field of artificial intelligence technology. The method includes: obtaining a training text dataset based on historical customer service voice data; performing quantum measurement processing on the target statement text training data in the training text dataset to obtain the semantic features of the target statement; using a bidirectional long short-term memory (BiLSTM) model and a self-attention mechanism to obtain the semantic features of the target dialogue based on the semantic features of the target statement; and obtaining a customer service outcome prediction model based on the semantic features of the target dialogue and the corresponding historical customer service outcomes. This invention's solution, by obtaining a customer service outcome prediction model based on the semantic features between the contexts of the target dialogue, fully models the semantic features of the target statement and the semantic features between the contexts of the target dialogue, increasing the accuracy of the customer service outcome prediction model in predicting customer service outcomes.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus, and device for establishing a customer service outcome prediction model. Background Technology

[0002] Telecommunications operators have accumulated a wealth of customer service data, such as customer service voice messages, outbound call voice messages, complaint tickets, online service records, and service tags. However, there is currently a lack of effective utilization in mining the value of this complaint data. Hotline customer service call voice data contains rich semantic information, yet this information is not being effectively used. By mining the rich semantic information in complaint voice data, potential customer dissatisfaction or complaint tendencies can be identified in advance, allowing operators to communicate with and appease potential complainants in a timely manner, thereby continuously improving the service experience.

[0003] Existing methods have been used to develop a series of user complaint prediction models (customer service outcome prediction models), but the following problems still exist: First, they do not fully integrate complaint scenarios and utilize data. For example, most current user complaint prediction models focus on user feature engineering while ignoring hotline voice data. Voice data contains user requests, user satisfaction levels, and complaint tendencies, and contains rich semantic information. Therefore, hotline voice data is valuable for studying users' potential complaint tendencies and can help improve the accuracy and reliability of prediction results. Second, existing complaint prediction models lack contextual semantic modeling capabilities, and their performance has room for improvement. Although some research has been conducted on user complaint prediction models in non-telecommunications fields using text data, the models used do not mine the semantics of words in context, i.e., they do not fully model the semantic interactions between words. However, according to existing research, the contextual semantics of words often helps in understanding text content. In other words, existing service prediction models do not consider the semantic features in voice data when modeling, leading to inaccurate predictions of service outcomes. Summary of the Invention

[0004] This invention provides a method, apparatus, and device for establishing a customer service outcome prediction model, which solves the problem in the prior art where the semantic features in voice data are not considered when modeling customer service outcome prediction models, resulting in inaccurate predictions of customer service outcomes.

[0005] To address the aforementioned technical problems, the embodiments of the present invention provide the following technical solutions:

[0006] This invention provides a method for establishing a customer service outcome prediction model, comprising:

[0007] Based on historical customer service voice data, a training text data set is obtained;

[0008] Quantum measurement processing is performed on the target sentence text training data in the training text dataset to obtain the semantic features of the target sentence;

[0009] Using a bidirectional long short-term memory (BiLSTM) model and a self-attention mechanism, the semantic features of the target dialogue are obtained based on the semantic features of the target sentence.

[0010] Based on the semantic features of the target dialogue and the historical customer service results corresponding to the target dialogue, a customer service result prediction model is obtained;

[0011] The target dialogue is any dialogue in the training text data set;

[0012] The target statement is any statement in the target dialogue.

[0013] Optionally, the method further includes:

[0014] Using the customer service result prediction model, the customer service result corresponding to the voice data to be predicted is obtained based on the voice data to be predicted.

[0015] Optionally, the step of performing quantum measurement processing on the target sentence text training data in the training text dataset to obtain the semantic features of the target sentence includes:

[0016] The target sentence text data is subjected to word vector embedding processing to obtain the word vectors of the words in the target sentence;

[0017] Using the BiLSTM model, the word vectors of the words in the target sentence are processed to obtain the hidden state vectors of the words in the target sentence;

[0018] The hidden state vectors of the words in the target sentence are encoded to obtain the pseudo-density matrix of the target sentence;

[0019] The semantic features of the target statement are obtained by projecting the pseudo-density matrix of the target statement.

[0020] Optionally, the step of using a BiLSTM model to process the word vectors of the words in the target sentence to obtain the hidden state vectors of the words in the target sentence includes:

[0021] Input the target statement into the BiLSTM model;

[0022] Obtain the hidden state vectors of the words in the target sentence output by the BiLSTM model;

[0023] The number of hidden layers in the BiLSTM model is the same as the number of words in the target sentence.

[0024] Optionally, encoding the hidden state vectors of words in the target sentence to obtain the quasi-density matrix of the target sentence includes:

[0025] The hidden state vectors of the words in the target sentence are normalized to obtain the quantum state vectors of the words in the target sentence;

[0026] The outer product of the quantum state vectors of the words in the target sentence is calculated to obtain the density matrix of the words in the target sentence;

[0027] The pseudo-density matrix of the target sentence is obtained based on the density matrix of words in the target sentence and the weight coefficient corresponding to each word in the target sentence.

[0028] Optionally, the step of projecting the pseudo-density matrix of the target statement to obtain the semantic features of the target statement includes:

[0029] The semantic features of the target statement are obtained by performing a preset number of projection calculations on the pseudo-density matrix of the target statement on the projection operator.

[0030] Optionally, the step of using a bidirectional long short-term memory (BiLSTM) model and a self-attention mechanism to obtain the semantic features of the target dialogue based on the semantic features of the target statement includes:

[0031] Using the BiLSTM model, the semantic features of multiple target statements in the target dialogue are processed to obtain the semantic features between the target statements in the target dialogue; the number of hidden layers in the BiLSTM model is equal to the number of target statements in the target dialogue.

[0032] By using a self-attention mechanism, the semantic features between the target statements are fused to obtain the abstract semantic features of the target dialogue.

[0033] This invention also provides a customer service outcome prediction model building apparatus, comprising:

[0034] The first processing module is used to obtain a training text data set based on historical customer service voice data;

[0035] The second processing module is used to perform quantum measurement processing on the target sentence text training data in the training text dataset to obtain the semantic features of the target sentence.

[0036] The third processing module is used to obtain the semantic features of the target dialogue based on the semantic features of the target sentence by utilizing the bidirectional long short-term memory (BiLSTM) model and the self-attention mechanism.

[0037] The model generation module is used to obtain a customer service result prediction model based on the semantic features of the target dialogue and the historical customer service results corresponding to the target dialogue.

[0038] The target dialogue is any dialogue in the training text data set;

[0039] The target statement is any statement in the target dialogue.

[0040] Optionally, the device further includes:

[0041] The customer service result generation module is used to obtain the customer service result corresponding to the voice data to be predicted based on the customer service result prediction model.

[0042] Optionally, the second processing module includes:

[0043] The first processing unit is used to perform word vector embedding processing on the target sentence text data to obtain word vectors of words in the target sentence;

[0044] The second processing unit is used to process the word vectors of the words in the target sentence using the BiLSTM model to obtain the hidden state vectors of the words in the target sentence.

[0045] The encoding unit is used to encode the hidden state vectors of words in the target sentence to obtain the pseudo-density matrix of the target sentence;

[0046] The measurement unit is used to perform projection measurement on the pseudo-density matrix of the target statement to obtain the semantic features of the target statement.

[0047] Optionally, the second processing unit is specifically used for:

[0048] Input the target statement into the BiLSTM model;

[0049] Obtain the hidden state vectors of the words in the target sentence output by the BiLSTM model;

[0050] The number of hidden layers in the BiLSTM model is the same as the number of words in the target sentence.

[0051] Optionally, the encoding unit is specifically used for:

[0052] The hidden state vectors of the words in the target sentence are normalized to obtain the quantum state vectors of the words in the target sentence;

[0053] The outer product of the quantum state vectors of the words in the target sentence is calculated to obtain the density matrix of the words in the target sentence;

[0054] The pseudo-density matrix of the target sentence is obtained based on the density matrix of words in the target sentence and the weight coefficient corresponding to each word in the target sentence.

[0055] Optionally, the measuring unit is specifically used for:

[0056] The semantic features of the target statement are obtained by performing a preset number of projection calculations on the pseudo-density matrix of the target statement on the projection operator.

[0057] Optionally, the third processing module includes:

[0058] The third processing unit is used to process the semantic features of multiple target statements in the target dialogue using the BiLSTM model to obtain the semantic features between the target statements in the target dialogue; the number of hidden layers in the BiLSTM model is equal to the number of target statements in the target dialogue.

[0059] The fusion unit is used to fuse the semantic features between the target statements through a self-attention mechanism to obtain the abstract semantic features of the target dialogue.

[0060] This invention also provides a customer service outcome prediction model building device, comprising: a processor, a memory, and a program stored in the memory and executable on the processor, wherein the program, when executed by the processor, implements the steps of the customer service outcome prediction model building method as described above.

[0061] This invention also provides a readable storage medium storing a program that, when executed by a processor, implements the steps in the customer service result prediction model establishment method described above.

[0062] The beneficial effects of this invention are:

[0063] This invention provides a training text dataset based on historical customer service voice data. Quantum measurement processing is applied to the target statement training data in the dataset to obtain the semantic features of the target statement. A Bidirectional Long Short-Term Memory (BiLSTM) model and a self-attention mechanism are used to obtain the semantic features of the target dialogue based on these features. Finally, a customer service result prediction model is derived based on the semantic features of the target dialogue and the corresponding historical customer service results. In essence, during the establishment of the customer service result prediction model, quantum measurement processing is used to obtain the semantic features of the dialogue statement (target statement), and the BiLSTM model and self-attention mechanism are used to obtain the semantic features between the contexts in the target dialogue. The customer service result prediction model obtained based on these contexts fully models the semantic features of the target statement and the semantic features between the contexts in the target dialogue, thus increasing the accuracy of the model in predicting customer service results. Attached Figure Description

[0064] Figure 1 A flowchart illustrating the customer service result prediction model establishment method provided in this embodiment of the invention;

[0065] Figure 2 This diagram illustrates the architecture of the customer service result prediction model provided in this embodiment of the invention.

[0066] Figure 3 This diagram illustrates the pseudo-density encoding and projection measurement of a target statement provided in an embodiment of the present invention.

[0067] Figure 4 This is a schematic diagram of the structure of the customer service result prediction model building device provided in an embodiment of the present invention;

[0068] Figure 5 This is a schematic diagram of the structure of the customer service result prediction model building device provided in this embodiment of the invention. Detailed Implementation

[0069] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0070] Before describing specific embodiments, the following will be explained first:

[0071] Non-classical probability

[0072] Quantum probability (non-classical probability) is a generalized probabilistic framework established by John von Neumann, aiming to explain quantum mechanics using mathematical language based on linear algebra. Besides describing physical particles in the microscopic world, quantum probability can also describe quantum-like phenomena in the macroscopic world, such as cognitive interference and the order effect in decision-making. Therefore, quantum probability can be naturally applied as a mathematical formalization method to fields such as human cognition, information retrieval, and natural language processing.

[0073] Events (also known as random events) are the cornerstone of probability theory and can be considered a manifestation of random phenomena. Classical probability corresponds to a set, with events being a subset of that set. The unordered nature of the set means that the order in which events occur does not affect the outcome. As for quantum probability, the quantum probability space is a Hilbert space (a generalization of Euclidean space, a complete, infinite-dimensional inner product space with the concepts of distance and angle). Each event corresponds to a subspace (such as a line, plane, or hyperplane), and different orders of occurrence may lead to different outcomes.

[0074] Quantum states and superposition states

[0075] Before introducing quantum states, let's first introduce the Dirac notation used to describe quantum mechanics: · represents the left vector, i.e., a row vector of unit length, and |·> represents the right vector, i.e., a column vector of unit length. A quantum state describes the state of a quantum system and can be viewed as a fundamental event in the quantum probability space, which can be represented by a unit column vector |μ>∈H. n This is represented as |μ>=(μ1,μ2,...,μ n The inner product of two vectors |μ> and |ν> is denoted as <μ|ν>.

[0076]

[0077] The cross product is denoted as |μ><ν|.

[0078]

[0079] A superposition state is a normalized linear combination of several quantum states. It is a key fundamental concept in quantum mechanics, describing the motion and uncertainty of the state of matter in the microscopic world. Extending this to the macroscopic world, various uncertainties can be considered to exist in a superposition state, such as Schrödinger's cat. Let's take the simplest quantum mechanical system, the single-qubit system, as an example. A single-qubit system has a two-dimensional state space. Assuming |0> and |1> are mutually orthogonal bases in this space, i.e., eigenstates, then any superposition state in this space can be represented as:

[0080] |φ>=α|0>+β|1>

[0081] Where α and β are complex numbers, called amplitudes (also known as probability amplitudes), satisfying |α| 2 +|β| 2 =1.

[0082] density matrix

[0083] Density matrices (also known as density operators) are another formal mathematical language. Mathematically equivalent to state vectors, they are more convenient for considering some general scenarios in quantum mechanics, one important one being the description of mixed states. A brief introduction to density matrices follows.

[0084] Suppose a quantum system with p i The probability is in state |φ i Then the density matrix of the system is defined as:

[0085]

[0086] Where, p i Let represent a non-negative probability value, and satisfy the condition that the sum of the probability values ​​is 1, i.e.:

[0087]

[0088] Furthermore, the diagonal elements of the density matrix can represent a valid probability distribution, while the off-diagonal elements represent the correlations of different dimensions of the quantum state. The properties of the density matrix are summarized as follows: (1) The density matrix is ​​a Hermitian matrix; (2) The trace of the density matrix is ​​1; (3) The density matrix is ​​a positive semi-definite matrix.

[0089] Quantum measurement

[0090] Quantum measurement allows us to observe quantum systems to understand their internal workings, using a set of measurement operators {M}. m The description states that the subscript m represents the possible measurement result in the measurement experiment. If the state of a quantum system being measured before the measurement is represented by ρ, then the probability p(m)∈[0,1] of producing the measurement result m is calculated by the following formula:

[0091]

[0092] Where * denotes the conjugate transpose, and the measurement operator satisfies the completeness relation, i.e.:

[0093]

[0094] The completeness relation expresses the inherent requirement that the sum of the measurement probabilities is 1, that is:

[0095]

[0096] Projection operators are a special type of measurement operator. While satisfying completeness, projection operators are mutually orthogonal. Measurements performed using projection operators are called projection measurements. The process of projection measurement can be viewed as the state vector |φ1> and |φ1> in the basis vector |ψ m The projection along the direction, with lengths p1 and p2, can represent the probability values ​​of obtaining the corresponding measurement result m.

[0097] pseudo-density matrix

[0098] By relaxing the constraint that the density matrix is ​​positive semi-definite, a quasi-density matrix can be obtained. The quasi-density matrix does not satisfy positive semi-definiteness, but inherits other properties of the density matrix: it satisfies the property that the trace is 1 and it is a Hermitian matrix. One of the biggest differences between the quasi-density matrix and the density matrix lies in the results of quantum measurements: the measurement result of a strictly defined density matrix is ​​always non-negative, while the measurement result of a quasi-density matrix may be negative. The model uses a formal method for quasi-density matrices, employing the partial transpose (PT) method to construct the quasi-density matrix. An example of the partial transpose of a 4x4 density matrix is ​​shown in Table 1 (the 2x2 matrices in the upper right and lower left corners of Table 1 undergo conjugate transpose transformations). The quasi-density matrix does not satisfy positive semi-definiteness, and the measurement result may be negative. In a sense, allowing the existence of negative values ​​gives the model a more flexible mathematical form and stronger information representation capabilities. For example, if the result obtained through quantum measurement in the model is $[-0.6, 0.4, 0.5]$, and the output obtained by probability normalization in the softmax function is $[0.15, 0.40, 0.45]$, then the order of the absolute values ​​of the data has obviously changed. However, when all the measurement results are positive, probability normalization does not produce a change in the order of the absolute values. Therefore, it can be considered that the measurement results of the quasi-density matrix are more significantly dependent on the context, i.e., they exhibit strong contextuality.

[0099] Table 1. Examples of the deflection of the density matrix.

[0100]

[0101] This invention addresses the problem in existing technologies where service prediction models fail to consider semantic features in voice data, leading to inaccurate predictions of service outcomes. It provides a method, apparatus, and device for predicting customer service outcomes.

[0102] like Figure 1 As shown, this embodiment of the invention provides a customer service outcome prediction method, including:

[0103] Step 101: Obtain the training text data set based on historical customer service voice data.

[0104] In this step, historical customer service data includes hotline voice data, which is the voice data during communication between hotline customer service representatives and users. Hotline voice data mainly includes voice data of complaints (voice data of dissatisfaction) and voice data of non-complaints (voice data of satisfaction). Hotline voice data records user requests, user satisfaction levels, and complaint tendencies, and contains a wealth of information.

[0105] Specifically, the training text dataset includes positive and negative samples. Positive samples are voice data of complaints (voice data of dissatisfaction), and negative samples are voice data of non-complaints (voice data of satisfaction).

[0106] Based on historical customer service data, a training text dataset was obtained. This training text dataset mainly covers two specific business scenarios: home broadband and wireless networks, filling the gap in text complaint datasets within the telecommunications industry.

[0107] Step 102: Perform quantum measurement processing on the target sentence text training data in the training text dataset to obtain the semantic features of the target sentence.

[0108] In this step, quantum measurement processing is performed on the training text dataset to obtain the sentence-level semantic features of the target sentence corresponding to the target sentence text training data in the training text dataset.

[0109] The target statement is any statement in the target dialogue, which can be a question-and-answer dialogue. Therefore, the target statement can be a question or an answer in the question-and-answer dialogue.

[0110] Step 103: Using the Bidirectional Long Short-Term Memory (BiLSTM) model and self-attention mechanism, obtain the semantic features of the target dialogue based on the semantic features of the target sentence.

[0111] The target dialogue is any dialogue in the training text data set;

[0112] The target statement is any statement in the target dialogue.

[0113] After obtaining the sentence-level semantic features of the target statement, this step utilizes a Bidirectional Long Short-Term Memory (BiLSTM) model to mine semantic information of word-level "context" weights, and fully models the semantic features of sentence-level context in the target dialogue using a sentence-level self-attention mechanism. This fully utilizes the rich semantic information in the voice data during the communication process between hotline customer service and users, thereby improving the prediction accuracy of the service prediction model.

[0114] Step 104: Based on the semantic features of the target dialogue and the historical customer service results corresponding to the target dialogue, obtain the customer service result prediction model.

[0115] The historical customer service results include complaints (dissatisfaction) and non-complaints (satisfaction).

[0116] In this step, the semantic features of the sentence-level context in the target dialogue and the corresponding complaint (dissatisfaction) and non-complaint (satisfaction) customer service results are used to train a customer service result prediction model, which is used to indicate the correspondence between customer service results and voice data.

[0117] In this embodiment of the invention, the established customer service outcome prediction model fully models the semantic features of the target statement and the semantic features between the target dialogue context, thereby improving the accuracy of predicting customer service outcomes when using the customer service outcome prediction model in the future.

[0118] Optionally, the method further includes:

[0119] Using the customer service result prediction model, the customer service result corresponding to the voice data to be predicted is obtained based on the voice data to be predicted.

[0120] After training the customer service outcome prediction model, the voice data to be predicted is obtained, and the voice data to be predicted is input into the customer service outcome prediction model. The model outputs the customer service outcome corresponding to the voice data to be predicted. The customer service outcome can be a complaint (dissatisfaction) or a non-complaint (satisfaction). That is, it outputs whether the user corresponding to the voice data to be predicted has a potential tendency to complain, which helps the operation and service managers to proactively and efficiently handle user complaints.

[0121] Optionally, the step of obtaining the training text data set based on historical customer service voice data includes:

[0122] The historical customer service voice data is processed by speech-to-text conversion to obtain the training text data set.

[0123] In this embodiment, the architecture for establishing the service prediction model is as follows: Figure 2 As shown, the architecture, from bottom to top, includes a data layer, a preprocessing layer, a cognition layer, and an output layer. The data layer stores historical customer service voice data and corresponding complaint records. The preprocessing layer acquires voice data of user complaints and voice data of non-complaints, performs speech-to-text processing, and obtains a training text dataset. Specifically, the voice text of user complaint records in the training text dataset is obtained by retrieving all inbound hotline voice data through service request discrimination, identifying the voice data of complaints, and transcribing it into text. The voice text of non-user complaint records in the training text dataset is obtained by randomly selecting voice data from "business consultation" and "business inquiry" service requests and transcribing it into text.

[0124] Optionally, the step of performing quantum measurement processing on the target sentence text training data in the training text dataset to obtain the semantic features of the target sentence includes:

[0125] The target sentence text data is subjected to word vector embedding processing to obtain the word vectors of the words in the target sentence;

[0126] Using the BiLSTM model, the word vectors of the words in the target sentence are processed to obtain the hidden state vectors of the words in the target sentence;

[0127] The hidden state vectors of the words in the target sentence are encoded to obtain the pseudo-density matrix of the target sentence;

[0128] The semantic features of the target statement are obtained by projecting the pseudo-density matrix of the target statement.

[0129] Please continue reading. Figure 2 The preprocessing layer performs speech-to-text conversion on historical customer service voice data to obtain a training text dataset. This dataset undergoes data cleaning and Chinese word segmentation. Then, the "Embedding Vector" module in the preprocessing layer maps each word in each sentence (target sentence) of the training text dataset to a word vector of dimension d (d ≥ 1). For simplicity, assume a voice text dataset contains m preprocessed sentence texts, where each sentence S (target sentence) contains n (n ≥ 1) words. In the "Embedding Vector" module, pre-trained Chinese word vectors are used, and to better suit the telecommunications domain, the word vectors in the model are further trained. After mapping, each word is formalized into a word vector w according to its order of appearance in the text. i The word vector representations of the words in the target sentence S are obtained as S = [w1, w2, ..., w n ], where wi ∈R d .

[0130] To model the semantic features of words within the context of each sentence, please refer to [link / reference needed]. Figure 2 The cognitive layer performs word-level contextual modeling (model building) and model training on the word vectors of words in the target statement S. The target statement S is input into the BiLSTM module to obtain the hidden state vectors of the words in the target statement. Then, the hidden state vectors of the words in the target statement S are encoded using a quasi-density matrix to obtain the quasi-density matrix of the target statement S. The quasi-density matrix of each statement is projected from the high-dimensional space to the low-dimensional subspace through projection measurement to extract the sentence-level semantic features of the statements in the quasi-density matrix, that is, to obtain the sentence-level semantic features of the target statement. These sentence-level semantic features are used for the subsequent modeling of the customer service result prediction model.

[0131] Optionally, the step of using a BiLSTM model to process the word vectors of the words in the target sentence to obtain the hidden state vectors of the words in the target sentence includes:

[0132] Input the target statement into the BiLSTM model;

[0133] Obtain the hidden state vectors of the words in the target sentence output by the BiLSTM model;

[0134] The number of hidden layers in the BiLSTM model is the same as the number of words in the target sentence.

[0135] Specifically, to model the contextual semantics of words within each sentence, the target sentence S is input into a Bidirectional Long Short-Term Memory (BiLSTM) module. This module has a number of hidden layers equal to the number of words in the sentence. Then, the contextual representation of the words in sentence S is obtained based on the hidden states output by the hidden layers.

[0136]

[0137]

[0138] Among them, w i Word vectors, representing words The hidden state vector representing a word. It includes the contextual semantic features of the words in the sentence.

[0139] Optionally, encoding the hidden state vectors of words in the target sentence to obtain the quasi-density matrix of the target sentence includes:

[0140] The hidden state vectors of the words in the target sentence are normalized to obtain the quantum state vectors of the words in the target sentence;

[0141] The outer product of the quantum state vectors of the words in the target sentence is calculated to obtain the density matrix of the words in the target sentence;

[0142] The pseudo-density matrix of the target sentence is obtained based on the density matrix of words in the target sentence and the weight coefficient corresponding to each word in the target sentence.

[0143] It should be noted that the quasi-density operator is an important semantic analysis method in quantum language models. The quasi-density operator (also called the density matrix) is a matrix used to describe the properties of quantum mixed states.

[0144] At the cognitive level, the semantic features of the target statement S are primarily obtained through quantum theories such as quasi-density matrices and quantum measurement. First, to model the contextual semantic features (semantic features between words) of each word in the target statement S, the word vectors of the words in the target statement S are processed to obtain the hidden state vectors of the words in the target statement S. These hidden state vectors include the contextual semantic features of the words in the target statement S. Then, quasi-density matrix encoding is performed on the hidden state vectors of the words in the target statement S to obtain the quasi-density matrix of the target statement S. Using projection measurement, the quasi-density matrix of each statement is projected from a high-dimensional space to a low-dimensional subspace to extract the sentence-level semantic features of the statements in the quasi-density matrix, thus obtaining the sentence-level semantic features of the target statement. The words in the training text dataset are mapped to vectors in Hilbert space, which are then used to construct the customer service result prediction model.

[0145] The word-level semantic features obtained by word-level context modeling of the target sentence S are then summed with context weights through a vector outer product operation in the model to obtain a sentence-level density matrix, which is further converted into a quasi-density matrix. Please refer to [link to relevant documentation]. Figure 3 The semantic features at the word level of the target sentence are encoded using a quasi-density matrix encoding method, resulting in a quasi-density matrix of the target sentence. The density matrix obtained through the outer product operation encodes the second-order relevance of words, while the quasi-density matrix obtained through certain mathematical transformations models strong contextuality. Furthermore, the context weights also depend on other words in the context, enabling the modeling of contextual relationships between words. The specific process is as follows:

[0146] First, the hidden state of each word in the target sentence is normalized using the L2 method to obtain the form of the quantum state vector, i.e.

[0147]

[0148] in, Describing the L2 norm, It can represent the word vector w in the target sentence i If the target statement is in a superposition state, then the target statement is represented in superposition state as follows:

[0149]

[0150] Then, according to the definition of the density matrix, the state vector The corresponding density matrix can be obtained by calculating the outer product:

[0151]

[0152] The off-diagonal elements of the density matrix represent the second-order correlation of vectors, i.e., the information interaction within words. Finally, the density matrix representation of the target sentence is obtained by summing the weight coefficients of each word and the density matrices of the corresponding words.

[0153]

[0154]

[0155] The density matrix of the target sentence can encode the contextual semantic information of all words in the target sentence. If the density matrix is ​​subjected to a partial transpose (PT) transformation, a quasi-density matrix representation can be obtained:

[0156] ρ′=PT(ρ)

[0157] It is easy to see that the density matrix calculated through the above steps can encode contextual information: firstly, the hidden state itself contains contextual information within the sentence; secondly, the weight coefficients calculated by the nonlinear formula... It is also context-dependent; the weight coefficient of each word is related to the other words in the sentence. Finally, thanks to the fact that the quasi-density matrix may produce negative values ​​during measurement, the probability normalization process becomes more context-dependent, thus exhibiting strong contextuality. Moreover, the word vectors are continuously updated with gradient descent during training, so the weight coefficients are also automatically updated accordingly.

[0158] Optionally, the step of projecting the pseudo-density matrix of the target statement to obtain the semantic features of the target statement includes:

[0159] The semantic features of the target statement are obtained by performing a preset number of projection calculations on the pseudo-density matrix of the target statement on the projection operator.

[0160] Projective measurement, a special quantum measurement technique, can represent the projection of a high-dimensional semantic space onto a low-dimensional semantic subspace, which can be analogous to data dimensionality reduction. The model uses projective measurement to project the quasi-density matrix of each sentence from a high-dimensional space to a low-dimensional subspace to extract the sentence-level semantic features encoded by the quasi-density matrix. Please continue reading. Figure 3 That is, by projecting the pseudo-density matrix, features are extracted to obtain sentence-level semantic features.

[0161] The model is initialized with a set of orthogonal vectors, and a set of projection measurement operators is obtained through L2 normalization and outer product. in Next, the model will calculate the projection of the density matrix representing the sentence onto each projection measurement operator according to the definition of projection measurement. After t (preset number) projection calculations, the measurement results can be obtained:

[0162] r′ i =tr(P i ρ j ), 1≤i≤t, 1≤j≤m,

[0163] r j =[r′1,r′2,...,r′ t ]

[0164] Where the scalar r' i The projection values ​​of a sentence in different subspaces within a semantic space can be viewed as abstract representations of the sentence's overall semantic information across these subspaces. Similarly, from a practical perspective, this chapter selectively ignores the completeness and orthogonality requirements of the measurement operator to ensure that the projection operator updates with the gradient during the model's training phase. This allows the most suitable projection operator to be learned based on the semantic information of a specific sentence.

[0165] For sentences represented by m density matrices within the call text (target dialogue), i.e., m target sentences, the semantic features of each sentence can be obtained through projection measurement:

[0166] R = [r1, r2, ..., r m ]

[0167] This invention implements a feature extraction method based on quantum measurement. After reducing the dimensionality of high-dimensional information, the probability of a vector in the semantic space falling under different measurement operators is calculated. This method can be used to infer the probability of generating a complaint tendency after customer service and output the results.

[0168] Optionally, the step of using a bidirectional long short-term memory (BiLSTM) model and a self-attention mechanism to obtain the semantic features of the target dialogue based on the semantic features of the target statement includes:

[0169] Using the BiLSTM model, the semantic features of multiple target statements in the target dialogue are processed to obtain the semantic features between the target statements in the target dialogue; the number of hidden layers in the BiLSTM model is equal to the number of target statements in the target dialogue.

[0170] By using a self-attention mechanism, the semantic features between the target statements are fused to obtain the abstract semantic features of the target dialogue.

[0171] Specifically, a single-layer BiLSTM is used to obtain the contextual semantic features between sentences in the target dialogue. First, the semantic features of each sentence in the target dialogue obtained from the decoding layer are input into a single-layer BiLSTM module. The number of hidden layers in the module is equal to the number of sentences, and the dimension of the hidden layers is d. h This module is used to model the contextual interactions between sentences, and outputs the semantic features between each sentence in the target dialogue as follows:

[0172]

[0173]

[0174] At this point, the model has completed the modeling of the contextual semantic interactions within and between sentences in the current call (target dialogue) text.

[0175] After obtaining the semantic features between sentences in the target dialogue, the semantic features of all sentences in the target dialogue are fused using a self-attention mechanism to obtain:

[0176]

[0177] Where f is an abstract representation (abstract semantic features) of the contextual semantic features of all sentences in the target dialogue, and it is input into a system such as... Figure 2 The fully connected layer (output layer) shown is trained based on the customer service result corresponding to the target dialogue, that is, whether the sample of the target dialogue is a positive sample or a negative sample, to obtain the customer service result prediction model.

[0178] To verify the model's performance, experiments were conducted on a speech-to-text dataset (the training text dataset). In application, the model was specifically designed for two service scenarios: wireless networks and home broadband. The non-classical probabilistic model was compared with a BiLSTM model combined with a self-attention mechanism. The wireless network dataset contained 5300 data points without complaints and 2125 data points with complaints; the home broadband dataset contained 3500 data points without complaints and 1409 data points with complaints.

[0179] The experimental results for the wireless network dataset are shown in Table 2. On the wireless network dataset, the non-classical probabilistic model achieves an accuracy of 0.9292, a precision of 0.8812, and a recall of 0.8683, surpassing the BiLSTM model combined with a self-attention mechanism. The experimental results for the home broadband dataset are shown in Table 3. The non-classical probabilistic model achieves a complaint prediction accuracy of 0.9844, a precision of 0.9615, and a recall of 0.9843, again surpassing the BiLSTM model combined with a self-attention mechanism.

[0180] Table 2 Experimental Results of Wireless Network Dataset

[0181]

[0182] Table 3 Experimental Results of the Home Broadband Dataset

[0183]

[0184] The system architecture of the non-classical probabilistic complaint prediction model includes modules such as data processing files, data files, data reading files, parameter configuration files, model definition files, and model inference files. The data preprocessing file is responsible for data preprocessing and generating data files; the data reading file is responsible for batch loading data from files; the parameter configuration file sets the main hyperparameters of the model; the model definition file defines the model class and its specific functions; and the main function file is responsible for data input and result output, coordinating the components of the model framework to perform their respective functions and jointly achieve model training and validation.

[0185] This invention provides a service prediction model based on non-classical probabilistics, which runs on the intelligent model layer of the technical architecture. Specifically, quantum model training is implemented on 26 training GPU servers in the AI ​​cluster, and the inference process for predicting user complaints is implemented on 20 inference GPU servers in the AI ​​cluster. The prediction results are generated in file format and connected to the Jingfen Big Data System. The prediction results are output to the Jingfen unified interface and then retrieved by CRM / BOSS / electronic channels through the interface to achieve precise operation and service.

[0186] The service prediction model constructed in this invention is based on quantum theory. It treats each word as a fundamental event in a quantum probability (non-classical probability) space, and sentences composed of words as a superposition of quantum fundamental events, represented by a quasi-density matrix capable of encoding second-order correlation and strong contextuality. Using a projection operator, information contained in the speech-to-text data is extracted through projective measurements. Furthermore, a Bidirectional Long Short-Term Memory (BiLSTM) network module, word-level "contextual weights," and sentence-level self-attention mechanisms are applied to fully model the contextual semantics at both the word and sentence levels. Supported by quantum theory, the service prediction model provided in this invention is more complex and effective.

[0187] The customer service outcome prediction method provided in this invention proposes a non-classical probability-based telecommunications complaint prediction model based on quantum theory. Under the non-classical probability framework, this model encodes semantic information using a quasi-density matrix and extracts semantic features through projective measurements. By combining a BiLSTM module and an attention mechanism, it models intra-word interactions, word-level contextual interactions, and sentence-level contextual interactions, providing support for modeling user satisfaction for accurate complaint prediction.

[0188] In constructing the service prediction model, this invention first maps each word in the call text to a word vector and inputs it into a word-level BiLSTM module for contextual interaction between words. Then, the model replaces the vector form with a quasi-density matrix and encodes the semantic information of the sentence by combining context weights. Subsequently, the model extracts the semantic information encoded by the quasi-density matrix by applying projection measurement technology to obtain sentence-level semantic features. Then, the model performs contextual interaction between sentences through a sentence-level BiLSTM and fuses sentence context features by combining a self-attention mechanism. Finally, the extracted features are input into a fully connected layer for user complaint or satisfaction classification to obtain the final output result.

[0189] This proposal presents a non-classical probabilistic complaint prediction method (customer service outcome prediction method) inspired by quantum theory. This method successfully implements a novel speech-to-text representation approach, mapping words in the speech-to-text to vectors in a semantic Hilbert space, which are then used for complaint text modeling. Simultaneously, it implements a feature extraction method based on quantum measurement, reducing the dimensionality of high-dimensional information and calculating the probability that vectors in the semantic space fall under different measurement operators. This probability can be used to infer the likelihood of a customer developing a complaint, and the results are then output.

[0190] The system architecture of the non-classical probabilistic complaint prediction model includes modules such as data processing files, data files, data reading files, parameter configuration files, model definition files, and model inference files. The data preprocessing file is responsible for data preprocessing and generating data files; the data reading file is responsible for batch loading data from files; the parameter configuration file sets the main hyperparameters of the model; the model definition file defines the model class and its specific functions; and the main function file is responsible for data input and result output, coordinating the components of the model framework to perform their respective functions and jointly achieve model training and validation.

[0191] This invention also proposes a quantum semantic feature representation method for hotline voice text of telecommunications operators. For hotline voice text and complaint text of telecommunications operators, each word is regarded as a basic event in quantum probability (non-classical probability) space, and the sentence composed of words is regarded as the superposition of quantum basic events. It is represented by a quasi-density matrix that can encode second-order correlation and strong contextuality. The hidden state of each word is normalized to obtain a quantum state vector. The quantum state vector is then multiplied by an outer product to obtain a density matrix. The context weight coefficients are summed to obtain the sentence density matrix. Projective measurement techniques are used to extract the semantic information encoded by the quasi-density matrix, thus obtaining sentence-level semantic features. A word-level "context weight" semantic information mining method using a Bidirectional Long Short-Term Memory (BiLSTM) network is proposed. To address the issue of traditional complaint prediction failing to fully utilize contextual semantic information, this method employs a BiLSTM module to achieve word-level "context weight" semantic information mining. A sentence-level self-attention mechanism is used to fully model word-level and sentence-level contextual semantics, fully utilizing the rich semantic information in the communication process between hotline customer service and users, thereby improving prediction accuracy. Finally, the contextual semantic features of all sentences are input into a fully connected layer / output layer, outputting the result indicating whether the user has a potential complaint tendency.

[0192] This solution effectively improves the interpretability of service prediction models. Traditional big data prediction models can only produce classification predictions of results, but cannot generate semantic reasons. This solution integrates non-classical probability models into the satisfaction system. Moreover, compared with other classifiers with weaker interpretability, this solution uses a knowledge semantic network to capture customers' potential intentions, which can uncover deeper intentions and achieve higher accuracy.

[0193] like Figure 4 As shown, this embodiment of the invention also provides a customer service result prediction model building device, including:

[0194] The first processing module 401 is used to obtain a training text data set based on historical customer service voice data;

[0195] The second processing module 402 is used to perform quantum measurement processing on the target sentence text training data in the training text data set to obtain the semantic features of the target sentence.

[0196] The third processing module 403 is used to obtain the semantic features of the target dialogue based on the semantic features of the target sentence by utilizing the bidirectional long short-term memory (BiLSTM) model and the self-attention mechanism.

[0197] The model generation module 404 is used to obtain a customer service result prediction model based on the semantic features of the target dialogue and the historical customer service results corresponding to the target dialogue.

[0198] The target dialogue is any dialogue in the training text data set;

[0199] The target statement is any statement in the target dialogue.

[0200] Optionally, the device further includes:

[0201] The customer service result generation module is used to obtain the customer service result corresponding to the voice data to be predicted based on the customer service result prediction model.

[0202] Optionally, the second processing module 402 includes:

[0203] The first processing unit is used to perform word vector embedding processing on the target sentence text data to obtain word vectors of words in the target sentence;

[0204] The second processing unit is used to process the word vectors of the words in the target sentence using the BiLSTM model to obtain the hidden state vectors of the words in the target sentence.

[0205] The encoding unit is used to encode the hidden state vectors of words in the target sentence to obtain the pseudo-density matrix of the target sentence;

[0206] The measurement unit is used to perform projection measurement on the pseudo-density matrix of the target statement to obtain the semantic features of the target statement.

[0207] Optionally, the second processing unit is specifically used for:

[0208] Input the target statement into the BiLSTM model;

[0209] Obtain the hidden state vectors of the words in the target sentence output by the BiLSTM model;

[0210] The number of hidden layers in the BiLSTM model is the same as the number of words in the target sentence.

[0211] Optionally, the encoding unit is specifically used for:

[0212] The hidden state vectors of the words in the target sentence are normalized to obtain the quantum state vectors of the words in the target sentence;

[0213] The outer product of the quantum state vectors of the words in the target sentence is calculated to obtain the density matrix of the words in the target sentence;

[0214] The pseudo-density matrix of the target sentence is obtained based on the density matrix of words in the target sentence and the weight coefficient corresponding to each word in the target sentence.

[0215] Optionally, the measuring unit is specifically used for:

[0216] The semantic features of the target statement are obtained by performing a preset number of projection calculations on the pseudo-density matrix of the target statement on the projection operator.

[0217] Optionally, the third processing module 403 includes:

[0218] The third processing unit is used to process the semantic features of multiple target statements in the target dialogue using the BiLSTM model to obtain the semantic features between the target statements in the target dialogue; the number of hidden layers in the BiLSTM model is equal to the number of target statements in the target dialogue.

[0219] The fusion unit is used to fuse the semantic features between the target statements through a self-attention mechanism to obtain the abstract semantic features of the target dialogue.

[0220] It should be noted that the customer service result prediction device provided in this embodiment of the invention is a device capable of executing the above-described customer service result prediction method. Therefore, all embodiments of the above-described customer service result prediction method are applicable to this device and can achieve the same or similar technical effects.

[0221] like Figure 5 As shown, this embodiment of the invention also provides a customer service result prediction model building device, including: a processor 500; and a memory 510 connected to the processor 500 via a bus interface, the memory 510 being used to store programs and data used by the processor 500 when performing operations, and the processor 500 calling and executing the programs and data stored in the memory 510.

[0222] The customer service result prediction model building device also includes a transceiver 520, which is connected to a bus interface and is used to receive and send data under the control of the processor 500.

[0223] Specifically, the processor 500 performs the following procedures:

[0224] Based on historical customer service voice data, a training text data set is obtained;

[0225] Quantum measurement processing is performed on the target sentence text training data in the training text dataset to obtain the semantic features of the target sentence;

[0226] Using a bidirectional long short-term memory (BiLSTM) model and a self-attention mechanism, the semantic features of the target dialogue are obtained based on the semantic features of the target sentence.

[0227] Based on the semantic features of the target dialogue and the historical customer service results corresponding to the target dialogue, a customer service result prediction model is obtained;

[0228] The target dialogue is any dialogue in the training text data set;

[0229] The target statement is any statement in the target dialogue.

[0230] Optionally, the processor 500 is further configured to:

[0231] Using the customer service result prediction model, the customer service result corresponding to the voice data to be predicted is obtained based on the voice data to be predicted.

[0232] Optionally, the processor 500 is configured to:

[0233] The target sentence text data is subjected to word vector embedding processing to obtain the word vectors of the words in the target sentence;

[0234] Using the BiLSTM model, the word vectors of the words in the target sentence are processed to obtain the hidden state vectors of the words in the target sentence;

[0235] The hidden state vectors of the words in the target sentence are encoded to obtain the pseudo-density matrix of the target sentence;

[0236] The semantic features of the target statement are obtained by projecting the pseudo-density matrix of the target statement.

[0237] Optionally, the processor 500 is specifically used for:

[0238] Input the target statement into the BiLSTM model;

[0239] Obtain the hidden state vectors of the words in the target sentence output by the BiLSTM model;

[0240] The number of hidden layers in the BiLSTM model is the same as the number of words in the target sentence.

[0241] Optionally, the processor 500 is specifically used for:

[0242] The hidden state vectors of the words in the target sentence are normalized to obtain the quantum state vectors of the words in the target sentence;

[0243] The outer product of the quantum state vectors of the words in the target sentence is calculated to obtain the density matrix of the words in the target sentence;

[0244] The pseudo-density matrix of the target sentence is obtained based on the density matrix of words in the target sentence and the weight coefficient corresponding to each word in the target sentence.

[0245] Optionally, the processor 500 is specifically used for:

[0246] The semantic features of the target statement are obtained by performing a preset number of projection calculations on the pseudo-density matrix of the target statement on the projection operator.

[0247] Optionally, the processor 500 is specifically used for:

[0248] Using the BiLSTM model, the semantic features of multiple target statements in the target dialogue are processed to obtain the semantic features between the target statements in the target dialogue; the number of hidden layers in the BiLSTM model is equal to the number of target statements in the target dialogue.

[0249] By using a self-attention mechanism, the semantic features between the target statements are fused to obtain the abstract semantic features of the target dialogue.

[0250] Among them, Figure 5 In this context, the bus architecture may include any number of interconnected buses and bridges, specifically linking various circuits together, represented by one or more processors (processor 500) and memory (memory 510). The bus architecture may also link together various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. A bus interface provides a user interface 530. A transceiver 520 may be multiple elements, including transmitters and receivers, providing units for communicating with various other devices over a transmission medium. Processor 500 is responsible for managing the bus architecture and general processing, and memory 510 may store data used by processor 500 during operation.

[0251] In addition, specific embodiments of the present invention also provide a readable storage medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the steps in the customer service outcome prediction method as described above.

[0252] In the several embodiments provided in this application, it should be understood that the disclosed methods and apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0253] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can be physically comprised separately, or two or more units can be integrated into one unit. The integrated unit described above can be implemented in hardware or in the form of hardware plus software functional units.

[0254] The integrated units implemented as software functional units described above can be stored in a computer-readable storage medium. These software functional units, stored in a storage medium, include several instructions that cause a computer device (which may be a personal computer, server, or network device, etc.) to execute some steps of the transmission and reception methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0255] The above describes the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also within the scope of protection of the present invention.

Claims

1. A method for establishing a customer service outcome prediction model, characterized in that, include: Based on historical customer service voice data, a training text data set is obtained; Quantum measurement processing is performed on the target sentence text training data in the training text dataset to obtain the semantic features of the target sentence; Using a bidirectional long short-term memory (BiLSTM) model and a self-attention mechanism, the semantic features of the target dialogue are obtained based on the semantic features of the target sentence; wherein, the target dialogue is any dialogue in the training text dataset; and the target sentence is any sentence in the target dialogue. Based on the semantic features of the target dialogue and the historical customer service results corresponding to the target dialogue, a customer service result prediction model is obtained; The step of performing quantum measurement processing on the target sentence text training data in the training text dataset to obtain the semantic features of the target sentence includes: The target sentence text training data is subjected to word vector embedding processing to obtain the word vectors of the words in the target sentence; Using the BiLSTM model, the word vectors of the words in the target sentence are processed to obtain the hidden state vectors of the words in the target sentence; The hidden state vectors of the words in the target sentence are encoded to obtain the pseudo-density matrix of the target sentence; The semantic features of the target statement are obtained by projecting the pseudo-density matrix of the target statement. The process of encoding the hidden state vectors of words in the target sentence to obtain the quasi-density matrix of the target sentence includes: The hidden state vectors of the words in the target sentence are normalized to obtain the quantum state vectors of the words in the target sentence; The outer product of the quantum state vectors of the words in the target sentence is calculated to obtain the density matrix of the words in the target sentence; The pseudo-density matrix of the target sentence is obtained based on the density matrix of words in the target sentence and the weight coefficient corresponding to each word in the target sentence.

2. The method for establishing a customer service outcome prediction model according to claim 1, characterized in that, The method further includes: Using the customer service result prediction model, the customer service result corresponding to the voice data to be predicted is obtained based on the voice data to be predicted.

3. The method for establishing a customer service outcome prediction model according to claim 1, characterized in that, The step of using a BiLSTM model to process the word vectors of words in the target sentence to obtain the hidden state vectors of the words in the target sentence includes: Input the target statement into the BiLSTM model; Obtain the hidden state vectors of the words in the target sentence output by the BiLSTM model; The number of hidden layers in the BiLSTM model is the same as the number of words in the target sentence.

4. The method for establishing a customer service outcome prediction model according to claim 1, characterized in that, The step of projecting the pseudo-density matrix of the target statement to obtain the semantic features of the target statement includes: The semantic features of the target statement are obtained by performing a preset number of projection calculations on the pseudo-density matrix of the target statement on the projection operator.

5. The method for establishing a customer service outcome prediction model according to claim 1, characterized in that, The method utilizes a bidirectional long short-term memory (BiLSTM) model and a self-attention mechanism to obtain the semantic features of the target dialogue based on the semantic features of the target statement, including: Using the BiLSTM model, the semantic features of multiple target statements in the target dialogue are processed to obtain the semantic features between the target statements in the target dialogue; the number of hidden layers in the BiLSTM model is equal to the number of target statements in the target dialogue. By using a self-attention mechanism, the semantic features between the target statements are fused to obtain the abstract semantic features of the target dialogue.

6. A customer service outcome prediction model building device, characterized in that, include: The first processing module is used to obtain a training text data set based on historical customer service voice data; The second processing module is used to perform quantum measurement processing on the target sentence text training data in the training text dataset to obtain the semantic features of the target sentence. The third processing module is used to obtain the semantic features of the target dialogue based on the semantic features of the target sentence by utilizing the bidirectional long short-term memory (BiLSTM) model and the self-attention mechanism; wherein, the target dialogue is any dialogue in the training text dataset; and the target sentence is any sentence in the target dialogue. The model generation module is used to obtain a customer service result prediction model based on the semantic features of the target dialogue and the historical customer service results corresponding to the target dialogue. The second processing module includes: The first processing unit is used to perform word vector embedding processing on the target sentence text training data to obtain word vectors of words in the target sentence; The second processing unit is used to process the word vectors of the words in the target sentence using the BiLSTM model to obtain the hidden state vectors of the words in the target sentence. The encoding unit is used to encode the hidden state vectors of words in the target sentence to obtain the pseudo-density matrix of the target sentence; A measurement unit is used to perform projection measurement on the pseudo-density matrix of the target statement to obtain the semantic features of the target statement; The encoding unit is specifically used for: The hidden state vectors of the words in the target sentence are normalized to obtain the quantum state vectors of the words in the target sentence; The outer product of the quantum state vectors of the words in the target sentence is calculated to obtain the density matrix of the words in the target sentence; The pseudo-density matrix of the target sentence is obtained based on the density matrix of words in the target sentence and the weight coefficient corresponding to each word in the target sentence.

7. A customer service outcome prediction model building device, characterized in that, include: A processor, a memory, and a program stored in the memory and executable on the processor, wherein the program, when executed by the processor, implements the steps of the customer service outcome prediction model building method as described in any one of claims 1 to 5.

8. A readable storage medium, characterized in that, The readable storage medium stores a program that, when executed by a processor, implements the steps in the customer service outcome prediction model building method as described in any one of claims 1 to 5.

Images