Online inquiry service answer quality false labeling method and system based on multi-source semantic feature fusion
By fusing multi-source semantic features and semi-supervised collaborative training, the problem of low accuracy in pseudo-annotation in online consultation services is solved, achieving high-precision pseudo-annotation results that are applicable to various online dialogue quality evaluation scenarios.
Patent Information
- Application Number
- CN202511862861.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-11
- Publication Date
- 2026-03-13
AI Technical Summary
In existing online consultation services, traditional pseudo-annotation methods rely on single semantic features and lack a systematic integration of medical knowledge features and global and local semantic interaction information. This results in a lot of noise in the pseudo-annotation results, affecting the generalization ability of the model. In addition, the data annotation cost is high, privacy is strong, and domain dependence is strong.
A multi-source semantic feature fusion method is adopted, which combines domain knowledge, global text semantics and local text semantics through a cross-feature fusion neural network. High-quality pseudo-annotation is achieved by using semi-supervised collaborative training and adaptive factorization network to achieve high-precision pseudo-annotation of online consultation service data.
It enhances the semantic expressiveness of pseudo-labels and the generalization ability of models, provides high-precision and interpretable pseudo-label results, and alleviates the problem of scarce data labels. It is applicable to various online dialogue quality evaluation scenarios such as online medical consultation, educational consultation, financial Q&A, and intelligent customer service.
Smart Images
Figure CN121660019A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the interdisciplinary field of artificial intelligence and medical informatics, specifically relating to a pseudo-labeling method and system for online consultation service answer quality based on multi-source semantic feature fusion. Background Technology
[0002] With the rapid development of internet-based healthcare services, online consultation platforms have become an important channel for patients to obtain health advice and medical guidance. The quality of doctors' answers directly affects the efficiency of doctor-patient interaction and the patient's treatment outcome. However, in practical applications, due to the non-standardized expression, semantic ambiguity, and mixed use of professional terminology in medical question-and-answer texts, traditional supervised learning-based answer quality prediction methods rely on large-scale manually labeled data. However, medical data labeling has problems such as high cost, strong privacy requirements, and strong domain dependence, resulting in scarce training data for models and limited model performance.
[0003] Existing semi-supervised pseudo-annotation methods mostly rely on single semantic features or non-textual features, such as word vector semantics or social features, lacking a systematic fusion of medical knowledge features and local and global semantic interaction information. Furthermore, traditional pseudo-annotation models often fail to fully consider the high-order nonlinear interactions between global and local semantics, resulting in noise in the pseudo-annotation results and affecting the model's generalization ability. Therefore, there is an urgent need for an innovative method and system that can integrate multi-source semantic information, combine domain knowledge, and achieve high-quality pseudo-annotation. Summary of the Invention
[0004] This invention aims to overcome the problems of low pseudo-annotation accuracy, insufficient semantic expression ability, and inadequate utilization of domain knowledge in existing technologies. It proposes a pseudo-annotation method and system for online consultation service answers based on multi-source semantic feature fusion, so as to achieve high-precision automatic pseudo-annotation of doctors' answer texts and provide reliable training data support for answer quality prediction models.
[0005] To achieve the above objectives, the present invention provides the following solution: A pseudo-annotation method for online medical consultation service answer quality based on multi-source semantic feature fusion includes: Collect and preprocess online consultation service data, extract features from the preprocessed online consultation service data, and obtain domain knowledge semantic features, global text semantic features, and local text semantic features; The domain knowledge semantic features are fused with global text semantic features and local text semantic features based on the cross-feature fusion neural network to obtain a global text semantic representation incorporating domain knowledge and a local text semantic representation incorporating domain knowledge. Semi-supervised collaborative training is performed on the global text semantic representation incorporating domain knowledge and the local text semantic representation incorporating domain knowledge. An ensemble learning strategy is adopted to obtain a joint high-order text semantic representation incorporating domain knowledge. By inputting a joint high-order text semantic representation incorporating domain knowledge into a factorization machine, pseudo-annotation results of the online consultation service answer quality are obtained.
[0006] Preferred methods for preprocessing online consultation service data include: The online consultation service data is anonymized for personal privacy and sensitive information, and the anonymized online consultation service data is cleaned, including removing HTML tags, emoticons, noisy characters and abnormal text; The cleaned online consultation service data is annotated by experts, and samples of different quality are randomly drawn in equal numbers according to category from the online consultation service data with expert annotation to obtain a class-balanced preprocessed online consultation service dataset. The preprocessed online consultation service dataset is divided into training and test sets according to a preset ratio.
[0007] Preferred methods for obtaining semantic features of domain knowledge include: By utilizing stacked convolutional neural network components with local context mechanisms, we learn local context semantic representations of preprocessed online consultation service data; By utilizing a long short-term memory network component with a global context mechanism, we learn the global context semantic representation of preprocessed online consultation service data; By utilizing a deep factorization machine component with a fusion context mechanism, the semantic features of the domain knowledge are obtained by fusing local context semantic representations and global context semantic representations.
[0008] Preferably, the method for extracting the global text semantic features includes: using a conductive long short-term memory network that adds the values of the input gate, forget gate, and memory unit to the loop path, combining the neighborhood information closest to the test set, capturing preset long-distance dependencies, and obtaining the global text semantic features.
[0009] Preferably, the method for obtaining the local text semantic features includes: using a semantically extended text convolutional neural network containing three core layers: a semantic extension layer, a multi-channel layer, and a multi-channel Text-CNN classifier, combined with external knowledge and capturing preset short-range dependencies at different levels and scales to obtain the local text semantic features.
[0010] Preferred methods for obtaining global text semantic representations incorporating domain knowledge include: Calculate the local self-attention weights and Kronecker product of the global text semantic features and the domain knowledge semantic features; The Kronecker product is weighted using the local self-attention weights and input into a fully connected layer containing only one neuron. The mutual preference weights are output through the activation function to construct a mutual preference matrix. The global text semantic features and domain knowledge semantic features are weighted based on the mutual preference matrix, and feature reconstruction is achieved through fully connected layers and residual connections to generate a global text semantic feature fusion stack and a domain knowledge semantic feature fusion stack. The Kronecker product self-attention mechanism is used to determine the relationship between the global text semantic feature fusion stack and the domain knowledge semantic feature fusion stack, and a relationship matrix is constructed. Based on the relation matrix, a global text semantic representation incorporating domain knowledge is obtained.
[0011] Preferred methods for obtaining joint high-order text semantic representations incorporating domain knowledge include: By inputting a global text semantic representation incorporating domain knowledge into a global adaptive factorization network, a global high-order text semantic representation incorporating domain knowledge is obtained. By inputting the local text semantic representation incorporating domain knowledge into the local adaptive factorization network, a local high-order text semantic representation incorporating domain knowledge is obtained. By taking the global high-order text semantic representation incorporating domain knowledge and the local high-order text semantic representation incorporating domain knowledge as two perspectives for collaborative training, a cross-pseudo-labeling strategy is adopted for iterative collaborative training. After a preset number of iterations, a joint adaptive factorization network is generated to obtain the joint high-order text semantic representation incorporating domain knowledge.
[0012] This invention also provides a pseudo-labeling system for the answer quality of online medical consultation services based on multi-source semantic feature fusion, used to implement the method, comprising: The multi-source semantic feature extraction module is used to collect and preprocess online consultation service data, and to extract features from the preprocessed online consultation service data to obtain domain knowledge semantic features, global text semantic features, and local text semantic features. The feature fusion module is used to fuse the domain knowledge semantic features with global text semantic features and local text semantic features respectively based on the cross-feature fusion neural network, so as to obtain a global text semantic representation incorporating domain knowledge and a local text semantic representation incorporating domain knowledge. The collaborative training module is used to perform semi-supervised collaborative training on the global text semantic representation and the local text semantic representation that incorporates domain knowledge. It adopts an ensemble learning strategy to obtain a joint high-order text semantic representation that incorporates domain knowledge. The pseudo-annotation module is used to input a joint high-order text semantic representation incorporating domain knowledge into a factorization machine to obtain pseudo-annotation results for the online consultation service answer quality.
[0013] This invention overcomes the bottlenecks in feature extraction and fusion of traditional pseudo-labeling methods for answer quality. Compared with existing technologies, this invention has the following advantages: Multi-source feature fusion enhances semantic expression capabilities: This invention comprehensively introduces three types of features: domain knowledge semantics, global text semantics, and local text semantics, breaking through the limitations of traditional single-source feature representation and significantly improving the completeness and depth of semantic information.
[0014] Knowledge injection mechanism based on cross-feature fusion neural network: Domain knowledge is injected at the global and local levels through cross-feature fusion network, so that the model has both domain specialization and language universality.
[0015] A semi-supervised collaborative training mechanism is introduced to alleviate the problem of scarce annotations: a global and local dual-view pseudo-annotation collaborative update mechanism is used to achieve high-quality self-learning of the model under the condition of a small amount of labeled data.
[0016] Introducing an adaptive factorization network to achieve high-order semantic modeling: By using global and local adaptive factorization networks to capture high-order interactive semantic relationships between different features within each feature perspective, the model's ability to model complex language phenomena is improved.
[0017] Accurate pseudo-labeling is achieved by combining a joint adaptive factorization network and factorization machine algorithm: Finally, a joint high-order text semantic representation incorporating domain knowledge is used as input, combined with the factorization machine classification algorithm from the recommender system domain, to obtain high-precision and highly interpretable pseudo-labeling results.
[0018] Highly scalable and versatile: This invention is not only applicable to the field of medical consultation, but can also be transferred to various online dialogue quality evaluation scenarios such as educational consultation, financial Q&A, and intelligent customer service. Attached Figure Description
[0019] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0020] Figure 1 This is a diagram illustrating the overall architecture of the method according to an embodiment of the present invention. Figure 2 This is a flowchart of the cross-feature fusion neural network algorithm according to an embodiment of the present invention. Detailed Implementation
[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0023] Example 1: like Figure 1 As shown, a pseudo-annotation method for online medical consultation service answer quality based on multi-source semantic feature fusion includes: S1: Collect and preprocess online consultation service data, extract features from the preprocessed online consultation service data to obtain domain knowledge semantic features, global text semantic features, and local text semantic features; a further implementation method includes: the method for preprocessing online consultation service data includes: The online consultation service data underwent anonymization to remove personal privacy and sensitive information. The anonymized data was then cleaned, including removing HTML tags, emoticons, noisy characters, and abnormal text. Specifically, the online consultation service dataset was collected from GitHub, and a combination of strong and weak anonymization methods was used to anonymize personal privacy and sensitive information (such as names, phone numbers, and ID numbers). The anonymized online consultation service data was then cleaned to ensure data quality.
[0024] The cleaned online consultation service data is annotated by experts. From this annotated data, samples of different quality are randomly selected in equal numbers according to category to obtain a class-balanced preprocessed online consultation service dataset. The preprocessed dataset is then divided into training and testing sets according to a preset ratio. Specifically, since the semi-supervised learning process involved in this invention also requires a small number of annotated samples to guide the learning of a large number of unannotated samples, three medical experts with rich clinical experience are invited to independently annotate each doctor's answer sample. The annotation results are aggregated using a majority voting mechanism, determining the final quality label (high quality or low quality) of the sample based on the principle of "majority rule," thereby reducing bias caused by individual subjectivity and improving the reliability and consistency of the annotation results.
[0025] To minimize bias during model training, this invention randomly extracts high-quality and low-quality samples in equal numbers according to category from a small number of quality-annotated response samples, thus forming a class-balanced preprocessed online consultation service dataset to effectively avoid model bias caused by class imbalance. Simultaneously, the preprocessed online consultation service dataset is divided into training and testing sets according to a preset ratio, maintaining consistency in class distribution across the subsets during the partitioning process to further ensure the fairness and reliability of performance evaluation during model training and testing.
[0026] A further implementation method for obtaining domain knowledge semantic features includes: This invention extracts domain knowledge semantic features from the output of Fusion_Context, a novel deep medical named entity recognition method based on fusion context mechanism (CSLBGDFM). CSLBGDFM can effectively alleviate the problem of out-of-vocabulary words by introducing ChineseBERT.
[0027] By utilizing stacked convolutional neural network components with local context mechanisms, we learn local context semantic representations of preprocessed online consultation service data; By utilizing a long short-term memory network component with a global context mechanism, we learn the global context semantic representation of preprocessed online consultation service data; By utilizing a deep factorization machine component with a fusion context mechanism, local context semantic representations and global context semantic representations are fused to obtain domain knowledge semantic features.
[0028] A further implementation method for extracting global text semantic features includes: employing a Transductive Long Short-Term Memory (T-LSTM) network that incorporates the values of the input gate, forget gate, and memory units into the recurrent path; combining this with neighborhood information closest to the test set; capturing preset long-distance dependencies; and obtaining global text semantic features. .
[0029] A further implementation method for obtaining local text semantic features includes: employing a semantically extended text convolutional neural network (SET-CNN) comprising three core layers: a semantic extension layer, a multi-channel layer, and a multi-channel Text-CNN classifier; combining external knowledge with the capture of preset short-range dependencies at different levels and scales to obtain local text semantic features. .
[0030] S2: The Cross-Feature Fusion Neural Network (CFFNN) fuses domain knowledge semantic features with global and local text semantic features to obtain global and local text semantic representations incorporating domain knowledge, thereby achieving text enhancement.
[0031] This invention constructs a cross-feature fusion neural network algorithm, taking the fusion of global text semantic features and domain knowledge semantic features as an example. Figure 2 As shown. A further implementation method for obtaining a global text semantic representation incorporating domain knowledge includes: Calculate the local self-attention weights of global text semantic features and domain knowledge semantic features, as well as the Kronecker product; weight the Kronecker product using the local self-attention weights, and input it into a fully connected layer containing only one neuron; output the mutual preference weights after passing through an activation function. A mutual preference matrix is constructed; based on the mutual preference matrix, the global text semantic features and the domain knowledge semantic features are weighted, and feature reconstruction is achieved through fully connected layers and residual connections to generate a global text semantic feature fusion stack and a domain knowledge semantic feature fusion stack; a Kronecker product self-attention mechanism is adopted to determine the relationship between the global text semantic feature fusion stack and the domain knowledge semantic feature fusion stack, and a relationship matrix is constructed; based on the relationship matrix, a global text semantic representation incorporating domain knowledge is obtained.
[0032] Specifically, starting with the global text semantic feature stack Given the first t eigenvectors From the domain knowledge semantic feature stack Given the first h eigenvectors Their local self-attention weights can be obtained according to Formula 1. : (1) in This represents the Kronecker product operation. The dimension representing the global text semantic feature vector or the domain knowledge semantic feature vector. This represents the activation function.
[0033] Then, the Kronecker product of each global text semantic feature vector and each domain knowledge semantic feature vector is calculated. Next, local self-attention is used to weight the Kronecker product. The resulting vector is then fed into a fully connected layer containing only one neuron, and finally passed through an activation function. The output is the mutual preference weights. As shown in Formula 2.
[0034] (2) in Represents the global text semantic feature stack No. t Each feature vector and domain knowledge semantic feature stack No. h The product of the weights of the eigenvectors This represents the product of the corresponding deviations. This represents the activation function. After obtaining all mutual preference weights... Construct a mutual preference matrix The next step in feature fusion is to use element-wise multiplication. This is what is executed. Specifically, the global text semantic feature stack. From the normalized mutual preference matrix Weighted, domain knowledge semantic feature stack Transposed normalized mutual preference matrix Weighting is performed as shown in Formula 3-6.
[0035] (3) (4) (5) (6) Cross-feature fusion neural networks employ a structure that creates a path from one hidden layer to all other hidden layers during stacking. Weighted domain knowledge semantic feature stack. By global text semantic feature stack Connection. Simultaneously, a weighted global text semantic feature stack. Domain knowledge semantic feature stack Connection. In this way, global textual semantics and domain knowledge semantic features are fused together. Using... The activation function projects the stack of connected vectors back to the network through a single-layer network. Matrices of the same dimension are then joined by residual connections to generate a global textual semantic feature fusion stack. As shown in Formula 7.
[0036] (7) in and Represents the global text semantic feature stack No. n The weights and biases of each feature vector. Similarly, the domain knowledge semantic feature fusion stack. It is implemented according to formula 8.
[0037] (8) in and Representation of the semantic feature stack of domain knowledge No. i The weights and biases of each feature vector. It's important to note that the parameters of the two fully connected branches are different.
[0038] Prediction Matrix The acquisition and mutual preference matrix Similarly, soon and Replace with and The modeling process is then performed again using Equations 1 and 2. Subsequently, this invention employs the Kronecker product self-attention mechanism to determine the relationship between global text semantic features and domain knowledge semantic features, obtaining a relationship matrix. As shown in Formula 9.
[0039] (9) in m Represents the global text semantic feature fusion stack or domain knowledge semantic feature fusion stack The dimension. Finally, this invention uses The activation function will convert the relation matrix The result is passed to a fully connected layer containing only one neuron, which then outputs the predicted score. matrix, This refers to the global text semantic representation that incorporates domain knowledge, as shown in Formula 10.
[0040] (10) in and Represents the global text semantic feature fusion stack or domain knowledge semantic feature fusion stack No. l The weights and biases of each eigenvector.
[0041] Similarly, the global text semantic feature stack in the above formula... Replace with local text semantic feature stack This allows us to obtain local textual semantic representations that incorporate domain knowledge. .
[0042] S3: Semi-supervised collaborative training is performed on the global text semantic representation and the local text semantic representation that incorporates domain knowledge. An ensemble learning strategy is adopted to obtain a joint high-order text semantic representation that incorporates domain knowledge. A further implementation method for obtaining a joint high-order text semantic representation incorporating domain knowledge includes: The global text semantic representation incorporating domain knowledge is input into a global adaptive factorization network to obtain a global high-order text semantic representation incorporating domain knowledge. The local text semantic representation incorporating domain knowledge is input into a local adaptive factorization network to obtain a local high-order text semantic representation incorporating domain knowledge. The global high-order text semantic representation and the local high-order text semantic representation incorporating domain knowledge are used as two perspectives for collaborative training. A cross-pseudo-labeling strategy is used for iterative collaborative training. After a preset number of iterations, a joint adaptive factorization network is generated to obtain a joint high-order text semantic representation incorporating domain knowledge.
[0043] This invention employs a semi-supervised collaborative training algorithm for pseudo-labeling and constructs a Global Adaptive Factorization Network (GAFN), a Local Adaptive Factorization Network (LAFN), and a Joint Adaptive Factorization Network (JAFN) based on the Adaptive Factorization Network (AFN) algorithm. These networks capture the high-order interactive semantic relationships between different features within the same feature perspective, representing global, local, and joint high-order textual semantic representations, respectively. The classifier uses the Factorization Machines (FM) algorithm, suitable for recommendation systems in feature-sparse scenarios. The pseudocode of the algorithm is shown in Table 1. Table 1
[0044] S4: Input the joint high-order text semantic representation incorporating domain knowledge into the factorization machine to obtain pseudo-annotation results of the online consultation service answer quality.
[0045] This invention selects precision, recall, and F1 score as performance metrics for model pseudo-labeling. Precision reflects the proportion of samples predicted as high-quality answers that are actually high-quality answers; recall represents the proportion of correctly identified samples among all samples that are actually high-quality answers. The F1 score comprehensively evaluates the overall performance of the method by using the harmonic mean of precision and recall. A higher F1 score indicates a better balance between precision and recall, especially in cases of imbalanced class distribution, where the F1 score more reasonably reflects the method's performance. The calculation methods for the above three metrics are as follows: , , .
[0046] Example 2 This invention also provides a pseudo-annotation system for the quality of online consultation service answers based on multi-source semantic feature fusion, used to implement the method of Embodiment 1, including: The multi-source semantic feature extraction module is used to collect and preprocess online consultation service data, and to extract features from the preprocessed online consultation service data to obtain domain knowledge semantic features, global text semantic features, and local text semantic features. The feature fusion module is used to fuse domain knowledge semantic features with global text semantic features and local text semantic features based on a cross-feature fusion neural network, so as to obtain a global text semantic representation incorporating domain knowledge and a local text semantic representation incorporating domain knowledge. The collaborative training module is used to perform semi-supervised collaborative training on the global text semantic representation and the local text semantic representation that incorporates domain knowledge. It adopts an ensemble learning strategy to obtain a joint high-order text semantic representation that incorporates domain knowledge. The pseudo-annotation module is used to input a joint high-order text semantic representation incorporating domain knowledge into a factorization machine to obtain pseudo-annotation results for the online consultation service answer quality.
[0047] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.
Claims
1. A pseudo-annotation method for online medical consultation service answer quality based on multi-source semantic feature fusion, characterized in that, include: Collect and preprocess online consultation service data, extract features from the preprocessed online consultation service data, and obtain domain knowledge semantic features, global text semantic features, and local text semantic features; The domain knowledge semantic features are fused with global text semantic features and local text semantic features based on the cross-feature fusion neural network to obtain a global text semantic representation incorporating domain knowledge and a local text semantic representation incorporating domain knowledge. Semi-supervised collaborative training is performed on the global text semantic representation incorporating domain knowledge and the local text semantic representation incorporating domain knowledge. An ensemble learning strategy is adopted to obtain a joint high-order text semantic representation incorporating domain knowledge. By inputting a joint high-order text semantic representation incorporating domain knowledge into a factorization machine, pseudo-annotation results of the online consultation service answer quality are obtained.
2. The method according to claim 1, characterized in that, Methods for preprocessing online consultation service data include: The online consultation service data is anonymized for personal privacy and sensitive information, and the anonymized online consultation service data is cleaned, including removing HTML tags, emoticons, noisy characters and abnormal text; The cleaned online consultation service data is annotated by experts, and samples of different quality are randomly drawn in equal numbers according to category from the online consultation service data with expert annotation to obtain a class-balanced preprocessed online consultation service dataset. The preprocessed online consultation service dataset is divided into training and test sets according to a preset ratio.
3. The method according to claim 1, characterized in that, Methods for obtaining semantic features of domain knowledge include: By utilizing stacked convolutional neural network components with local context mechanisms, we learn local context semantic representations of preprocessed online consultation service data; By utilizing a long short-term memory network component with a global context mechanism, we learn the global context semantic representation of preprocessed online consultation service data; By utilizing a deep factorization machine component with a fusion context mechanism, the semantic features of the domain knowledge are obtained by fusing local context semantic representations and global context semantic representations.
4. The method according to claim 2, characterized in that, The method for extracting the global text semantic features includes: using a conductive long short-term memory network that adds the values of the input gate, forget gate, and memory unit to the loop path, combining the neighborhood information closest to the test set, capturing preset long-distance dependencies, and obtaining the global text semantic features.
5. The method according to claim 2, characterized in that, The method for obtaining the local text semantic features includes: using a semantically extended text convolutional neural network containing three core layers: a semantic extension layer, a multi-channel layer, and a multi-channel Text-CNN classifier, combined with external knowledge and capturing preset short-distance dependencies at different levels and scales to obtain the local text semantic features.
6. The method according to claim 2, characterized in that, Methods for obtaining global text semantic representations incorporating domain knowledge include: Calculate the local self-attention weights and Kronecker product of the global text semantic features and the domain knowledge semantic features; The Kronecker product is weighted using the local self-attention weights and input into a fully connected layer containing only one neuron. The mutual preference weights are output through the activation function to construct a mutual preference matrix. The global text semantic features and domain knowledge semantic features are weighted based on the mutual preference matrix, and feature reconstruction is achieved through fully connected layers and residual connections to generate a global text semantic feature fusion stack and a domain knowledge semantic feature fusion stack. The Kronecker product self-attention mechanism is used to determine the relationship between the global text semantic feature fusion stack and the domain knowledge semantic feature fusion stack, and a relationship matrix is constructed. Based on the relation matrix, a global text semantic representation incorporating domain knowledge is obtained.
7. The method according to claim 2, characterized in that, Methods for obtaining joint high-order text semantic representations incorporating domain knowledge include: By inputting a global text semantic representation incorporating domain knowledge into a global adaptive factorization network, a global high-order text semantic representation incorporating domain knowledge is obtained. By inputting the local text semantic representation incorporating domain knowledge into the local adaptive factorization network, a local high-order text semantic representation incorporating domain knowledge is obtained. By taking the global high-order text semantic representation incorporating domain knowledge and the local high-order text semantic representation incorporating domain knowledge as two perspectives for collaborative training, a cross-pseudo-labeling strategy is adopted for iterative collaborative training. After a preset number of iterations, a joint adaptive factorization network is generated to obtain the joint high-order text semantic representation incorporating domain knowledge.
8. A pseudo-annotation system for online medical consultation service answer quality based on multi-source semantic feature fusion, used to implement the method described in any one of claims 1-7, characterized in that, include: The multi-source semantic feature extraction module is used to collect and preprocess online consultation service data, and to extract features from the preprocessed online consultation service data to obtain domain knowledge semantic features, global text semantic features, and local text semantic features. The feature fusion module is used to fuse the domain knowledge semantic features with global text semantic features and local text semantic features respectively based on the cross-feature fusion neural network, so as to obtain a global text semantic representation incorporating domain knowledge and a local text semantic representation incorporating domain knowledge. The collaborative training module is used to perform semi-supervised collaborative training on the global text semantic representation and the local text semantic representation that incorporates domain knowledge. It adopts an ensemble learning strategy to obtain a joint high-order text semantic representation that incorporates domain knowledge. The pseudo-annotation module is used to input a joint high-order text semantic representation incorporating domain knowledge into a factorization machine to obtain pseudo-annotation results for the online consultation service answer quality.