Large-scale language model question-answering system based on health-care knowledge

Abstract

The invention discloses a large-scale language model question-answering system based on health-care knowledge, which retrieves or reasones the most suitable answer from a health-care knowledge base by analyzing the natural language question of a user by utilizing an artificial intelligence technology, and returns the suitable answer to the user in a natural language form. Therefore, by applying a large-scale language model question-answering system in combination with health-care knowledge, accurate and comprehensive health-care question answering and guidance can be provided for the user, and the requirements of the user in the health-care field are met, so that the health consciousness and life quality of the user are improved, and auxiliary diagnosis and treatment suggestions can be provided for medical institutions and professionals.

Description

Technical Field

[0001] The present application relates to the field of language model question answering, and more specifically, to a large-scale language model question answering system based on health care knowledge. Background Art

[0002] Health care knowledge refers to relevant knowledge about health, health preservation, prevention and treatment of diseases, including traditional Chinese medicine, Western medicine, nutrition, exercise, psychology and other aspects. It is an important reference for people to improve their quality of life and health level.

[0003] As people pay more and more attention to health and quality of life, the knowledge and demand in the field of health care are also growing. However, the acquisition and application of health care knowledge face some challenges, such as knowledge dispersion, information redundancy, language barriers, etc. Specifically, the sources of health care knowledge are diverse, including scientific research, professional institutions, authoritative media, etc. However, information from different sources may differ or even contradict each other. When users acquire health care knowledge, they often face information fragmentation and inconsistency, and it is difficult to determine which information is credible. At the same time, the knowledge in the field of health care is updated rapidly and covers a wide range of content. When users acquire health care knowledge, they often face a large amount of information, which is difficult to effectively screen and organize. In addition, different information may conflict and be confusing, which brings confusion and uncertainty to users.

[0004] Therefore, a large-scale language model question-answering system based on health care knowledge is desired. Summary of the invention

[0005] In order to solve the above-mentioned technical problems, the present application is proposed. The embodiment of the present application provides a large-scale language model question-and-answer system based on health care knowledge, which uses artificial intelligence technology to analyze the user's natural language questions, retrieve or infer the most appropriate answers from the health care knowledge base, and return the appropriate answers to the user in the form of natural language. In this way, by applying a large-scale language model question-and-answer system combined with health care knowledge, users can be provided with accurate and comprehensive answers and guidance to health care questions, meet the needs of users in the field of health care, thereby improving users' health awareness and quality of life, and can also provide auxiliary diagnosis and treatment suggestions for medical institutions and professionals.

[0006] According to one aspect of the present application, a large-scale language model question-answering system based on health care knowledge is provided, which includes: a question collection module, used to obtain health care questions input by users; a question semantic supplement module, used to input the health care questions into a semantic compensator based on the ChatGPT model to obtain an optimized health care question description; a question semantic understanding module, used to perform semantic encoding processing on the optimized health care question description to obtain a sequence of health care question word granularity description feature vectors; an answer acquisition module, used to extract a first alternative health care answer from a background database; an answer semantic understanding module, used to perform semantic encoding processing on the first alternative health care answer to obtain a sequence of first alternative answer word granularity description feature vectors; a semantic interaction fusion analysis module, used to perform word granularity semantic sequence interaction fusion analysis on the sequence of health care question word granularity description feature vectors and the sequence of the first alternative answer word granularity description feature vectors to obtain question-answer word granularity semantic interaction fusion features; an answer return judgment module, used to determine whether to return the first alternative health care answer based on the question-answer word granularity semantic interaction fusion features.

[0007] Compared with the prior art, the present application provides a large-scale language model question-and-answer system based on health care knowledge, which uses artificial intelligence technology to analyze the user's natural language questions, retrieve or infer the most appropriate answer from the health care knowledge base, and return the appropriate answer to the user in the form of natural language. In this way, by applying a large-scale language model question-and-answer system combined with health care knowledge, users can be provided with accurate and comprehensive answers and guidance to health care questions, meet the needs of users in the field of health care, thereby improving users' health awareness and quality of life, and can also provide auxiliary diagnosis and treatment suggestions for medical institutions and professionals. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0009] Figure 1 This is a block diagram of a large-scale language model question and answer system based on health care knowledge according to an embodiment of the present application.

[0010] Figure 2 This is a system architecture diagram of a large-scale language model question and answer system based on health care knowledge according to an embodiment of the present application.

[0011] Figure 3This is a block diagram of the training phase of a large-scale language model question-answering system based on health care knowledge according to an embodiment of the present application.

[0012] Figure 4 This is a block diagram of an answer return judgment module in a large-scale language model question and answer system based on health care knowledge according to an embodiment of the present application. DETAILED DESCRIPTION

[0013] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0014] As shown in this application and claims, unless the context clearly indicates an exception, the words "a", "an", "an" and / or "the" do not refer to the singular and may also include the plural. Generally speaking, the terms "include" and "comprise" only indicate the inclusion of the steps and elements that have been clearly identified, and these steps and elements do not constitute an exclusive list. The method or device may also include other steps or elements.

[0015] Although the present application makes various references to certain modules in the system according to the embodiments of the present application, any number of different modules can be used and run on the user terminal and / or server. The modules are only illustrative, and different aspects of the system and method can use different modules.

[0016] Flowcharts are used in the present application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed accurately in order. On the contrary, various steps may be processed in reverse order or simultaneously as required. Meanwhile, other operations may also be added to these processes, or a certain step or several steps of operations may be removed from these processes.

[0017] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0018] As people pay more and more attention to health and quality of life, the knowledge and demand in the field of health care are also growing. However, the acquisition and application of health care knowledge face some challenges, such as knowledge dispersion, information redundancy, language barriers, etc. Specifically, the sources of health care knowledge are diverse, including scientific research, professional institutions, authoritative media, etc. However, information from different sources may differ or even contradict each other. When users acquire health care knowledge, they often face information fragmentation and inconsistency, and it is difficult to determine which information is credible. At the same time, the knowledge in the field of health care is updated rapidly and covers a wide range of content. When users acquire health care knowledge, they often face a large amount of information, which is difficult to effectively screen and organize. In addition, different information may conflict and be confusing, which brings confusion and uncertainty to users. Therefore, a large-scale language model question-answering system based on health care knowledge is expected.

[0019] In the technical solution of the present application, a large-scale language model question-answering system based on health care knowledge is proposed. Figure 1 This is a block diagram of a large-scale language model question and answer system based on health care knowledge according to an embodiment of the present application. Figure 2 : is a system architecture diagram of a large-scale language model question-answering system based on health care knowledge according to an embodiment of the present application. Figure 1 and Figure 2 As shown, according to the embodiment of the present application, the large-scale language model question-answering system 300 based on health care knowledge includes: a question collection module 310, which is used to obtain health care questions input by users; a question semantic supplement module 320, which is used to input the health care questions into a semantic compensator based on the ChatGPT model to obtain an optimized health care question description; a question semantic understanding module 330, which is used to perform semantic encoding processing on the optimized health care question description to obtain a sequence of health care question word granularity description feature vectors; an answer acquisition module 340, which is used to extract the first candidate health care answer from the background database The answer semantic understanding module 350 is used to perform semantic encoding processing on the first alternative health care answer to obtain a sequence of word granularity description feature vectors of the first alternative answer; the semantic interaction fusion analysis module 360 ​​is used to perform word granularity semantic sequence interaction fusion analysis on the sequence of word granularity description feature vectors of the health care question and the sequence of word granularity description feature vectors of the first alternative answer to obtain question-answer word granularity semantic interaction fusion features; the answer return judgment module 370 is used to determine whether to return the first alternative health care answer based on the question-answer word granularity semantic interaction fusion features.

[0020] In particular, the problem acquisition module 310 and the problem semantic supplement module 320 are used to obtain health care problems input by the user, and input the health care problems into the semantic compensator based on the ChatGPT model to obtain an optimized description of the health care problems. That is, first, the health care problems input by the user are obtained. It should be understood that the health care problems input by the user may be unclear, ambiguous or incomplete, which will affect the understanding of the problem and the accuracy of the answer. The ChatGPT model is a natural language processing model based on deep learning, with powerful semantic understanding and generation capabilities. Therefore, in order to facilitate a more accurate semantic understanding of the health care problems input by the user in the future, in the technical solution of the present application, the health care problems are further input into the semantic compensator based on the ChatGPT model to obtain an optimized description of the health care problems. By inputting the problem into the semantic compensator based on the ChatGPT model, the problem can be semantically compensated and optimized, making the description of the problem more accurate, clear and specific.

[0021] In particular, the problem semantic understanding module 330 is used to perform semantic encoding processing on the description of the problem of optimizing health care to obtain a sequence of feature vectors of word granularity description of health care problems. That is, in the technical solution of the present application, after the word segmentation processing of the description of the problem of optimizing health care, semantic encoding processing is performed in a problem semantic encoder including a word embedding layer to extract the contextual semantic association feature information about health care problems based on the global word granularity in the description of the problem of optimizing health care, thereby obtaining a sequence of feature vectors of word granularity description of health care problems. Specifically, the description of the problem in optimizing health care is segmented to convert the description of the problem in optimizing health care into a word sequence composed of multiple words; the embedding layer of the problem semantic encoder including the word embedding layer is used to map each word in the word sequence into a word embedding vector to obtain a sequence of word embedding vectors; the converter of the problem semantic encoder including the word embedding layer is used to perform global context semantic encoding based on the converter idea on the sequence of word embedding vectors to obtain multiple global context semantic feature vectors; and the multiple global context semantic feature vectors are cascaded to obtain a sequence of feature vectors of word granularity description of the problem in health care. More specifically, a converter of the question semantic encoder including a word embedding layer is used to perform global context semantic encoding based on the converter idea on the sequence of word embedding vectors to obtain multiple global context semantic feature vectors, including: arranging the sequence of word embedding vectors in one dimension to obtain a global feature vector; calculating the product between the global feature vector and the transpose vector of each word embedding vector in the sequence of word embedding vectors to obtain multiple self-attention association matrices; respectively normalizing each of the multiple self-attention association matrices to obtain multiple standardized self-attention association matrices; passing each of the multiple standardized self-attention association matrices through a Softmax classification function to obtain multiple probability values; respectively weighting the transpose vector of each word embedding vector in the sequence of word embedding vectors using each of the multiple probability values ​​as a weight to obtain the multiple context semantic feature vectors; and, cascading the multiple context semantic feature vectors to obtain the multiple global context semantic feature vectors.

[0022] In particular, the answer acquisition module 340 and the answer semantic understanding module 350 are used to extract the first alternative health care answer from the background database, and to perform semantic encoding processing on the first alternative health care answer to obtain a sequence of word granularity description feature vectors of the first alternative answer. That is, for the answer, first, the first alternative health care answer is extracted from the background database. Then, the first alternative health care answer is processed by word segmentation and semantic understanding is performed in the answer semantic encoder including a word embedding layer, so as to extract the contextual semantic association feature information based on the global word granularity in the first alternative health care answer, thereby obtaining a sequence of word granularity description feature vectors of the first alternative answer. Specifically, word segmentation is performed on the first alternative health care answer to convert the first alternative health care answer into a word sequence composed of multiple words; the embedding layer of the answer semantic encoder including the word embedding layer is used to map each word in the word sequence into a word embedding vector to obtain a sequence of word embedding vectors; a converter of the answer semantic encoder including the word embedding layer is used to perform global context semantic encoding based on the converter idea on the sequence of word embedding vectors to obtain multiple global context semantic feature vectors; and the multiple global context semantic feature vectors are cascaded to obtain a sequence of word granularity description feature vectors of the first alternative answer.

[0023] In particular, the semantic interaction fusion analysis module 360 ​​is used to perform word granularity semantic sequence interaction fusion analysis on the sequence of the word granularity description feature vectors of the health care aspect question and the sequence of the word granularity description feature vectors of the first alternative answer to obtain the question-answer word granularity semantic interaction fusion feature. Considering the complex semantic relationship between the health care aspect question and the first alternative health care aspect answer, in order to better understand the background of the question and the relevance of the answer, and improve the accuracy and relevance of the answer output, in the technical solution of the present application, the sequence of the word granularity description feature vectors of the health care aspect question and the sequence of the word granularity description feature vectors of the first alternative answer are further passed through the word granularity semantic sequence interaction fusion module to obtain the question-answer word granularity semantic interaction fusion feature vector. It should be understood that the word granularity semantic sequence interaction fusion module can weight the word granularity description feature vectors of each health care aspect question and the word granularity description feature vector of each first alternative answer by calculating the attention weight to model the importance between different feature sequences. Specifically, the module can assign weights to each feature sequence based on the semantic similarity and correlation between the word granularity semantic sequences of the sequence of feature vectors describing the health care questions at word granularity and the sequence of feature vectors describing the first alternative answers at word granularity. Therefore, after the attention interaction fusion processing, the sequence of feature vectors describing the health care questions at word granularity and the sequence of feature vectors describing the first alternative answers at word granularity can influence and complement each other through each local temporal feature vector in the word granularity semantic sequence, so as to more fully capture the interactive and collaborative feature information between the semantic features of the health care questions and the semantic features of the first alternative answers, thereby obtaining the question-answer word granularity semantic interaction fusion feature vector. This helps the system understand the meaning of the question more accurately, and match and compare it with the answer.Specifically, the sequence of the health care question word granularity description feature vectors and the sequence of the first alternative answer word granularity description feature vectors are passed through a word granularity semantic sequence interactive fusion module to obtain a question-answer word granularity semantic interactive fusion feature vector as the question-answer word granularity semantic interactive fusion feature, including: passing the sequence of the health care question word granularity description feature vectors through a health care question feature extractor based on a fully convolutional neural network model to obtain a sequence of health care question word granularity description full convolution feature vectors; passing the sequence of the first alternative answer word granularity description feature vectors Arrange as a global feature vector describing the word granularity of the first alternative answer; fuse the sequence of the fully convolutional feature vectors describing the word granularity of the health care question and the global feature vector describing the word granularity of the first alternative answer to obtain a question-answer word granularity semantic fusion feature vector; perform semantic association encoding on the sequence of the feature vectors describing the word granularity of the first alternative answer to obtain a global semantic association feature vector for the first alternative answer; fuse the question-answer word granularity semantic fusion feature vector and the first alternative answer global semantic association feature vector to obtain the question-answer word granularity semantic interaction fusion feature vector. More specifically, the sequence of the first candidate answer word granularity description feature vectors is semantically associated with the first candidate answer to obtain a global semantic association feature vector of the first candidate answer, including: one-dimensionally arranging the sequence of the first candidate answer word granularity description feature vectors to obtain a global feature vector of the first candidate answer; calculating the product of the first candidate answer global feature vector and the transposed vector of each first candidate answer word granularity description feature vector in the sequence of the first candidate answer word granularity description feature vector to obtain a plurality of self-attention association matrices; and performing standard on each of the plurality of self-attention association matrices. The invention further comprises the following steps: performing quantization processing to obtain a plurality of standardized self-attention association matrices; applying a Softmax classification function to each of the plurality of standardized self-attention association matrices to obtain a plurality of probability values; using each of the plurality of probability values ​​as a weight to weight each of the first alternative answer word granularity description feature vectors in the sequence of the first alternative answer word granularity description feature vectors to obtain the plurality of first alternative answer contextual semantic feature vectors; and concatenating the plurality of first alternative answer contextual semantic feature vectors to obtain the first alternative answer global semantic association feature vector.

[0024] It is worth noting that the Fully Convolutional Neural Network (FCN) is a special type of Convolutional Neural Network (CNN). Unlike traditional CNNs, the output of FCN is a dense pixel-level prediction, not just a classification or labeling of the entire image. The following are the main features and steps of the fully convolutional neural network: Convolutional layer: FCN is mainly composed of convolutional layers, which can extract image features through convolution operations. These convolutional layers can contain different filters to capture features of different scales and semantics; Upsampling: FCN uses upsampling operations to restore low-resolution feature maps to the resolution of the original image. Common upsampling methods include deconvolution and bilinear interpolation; Skip connection: In order to better preserve the spatial information of the image, FCN introduces skip connections. Skip connections connect low-level features with high-level features, allowing the network to simultaneously utilize feature information at different levels; Output layer: The output layer of FCN is a dense pixel-level prediction that classifies or labels each pixel. The output layer usually uses an appropriate activation function to generate the prediction probability of each pixel.

[0025] In particular, the answer return judgment module 370 is used to determine whether to return the first candidate health care answer based on the question-answer word granularity semantic interaction fusion feature. In particular, in a specific example of the present application, Figure 4 As shown, the answer return judgment module 370 includes: a fitness evaluation unit 371, which is used to pass the question-answer word granularity semantic interaction fusion feature vector through a classifier to obtain a classification result, and the classification result is used to indicate whether the fitness between the question and the answer exceeds a predetermined threshold; an answer return unit 372, which is used to determine whether to return the first alternative health care answer based on the classification result.

[0026] Specifically, the fitness evaluation unit 371 is used to pass the question-answer word granularity semantic interaction fusion feature vector through a classifier to obtain a classification result, and the classification result is used to indicate whether the fitness between the question and the answer exceeds a predetermined threshold. That is, the classification process is performed by using the semantic interaction fusion feature information between the semantic features of the health care question and the semantic features of the first alternative health care answer, so as to evaluate the fitness between the question and the answer, thereby judging whether the fitness exceeds a predetermined threshold, and determining whether to return the first alternative health care answer based on the result. In this way, users can acquire more health care knowledge, improve health awareness and quality of life, and can also provide auxiliary diagnosis and treatment suggestions for medical institutions and professionals. More specifically, the question-answer word granularity semantic interaction fusion feature vector is fully connected and encoded using multiple fully connected layers of the classifier to obtain an encoded classification feature vector; and the encoded classification feature vector is passed through the Softmax classification function of the classifier to obtain the classification result.

[0027] A classifier is a machine learning model or algorithm that is used to classify input data into different categories or labels. Classifiers are part of supervised learning and perform classification tasks by learning the mapping relationship from input data to output categories.

[0028] A fully connected layer is a common type of layer in neural networks. In a fully connected layer, each neuron is connected to all neurons in the previous layer, and each connection has a weight. This means that each neuron in the fully connected layer receives inputs from all neurons in the previous layer, and performs a weighted sum of these inputs through the weights, and then passes the result to the next layer.

[0029] The Softmax classification function is a commonly used activation function for multi-classification problems. It converts each element of the input vector into a probability value between 0 and 1, and the sum of these probability values ​​is equal to 1. The Softmax function is often used in the output layer of a neural network, especially for multi-classification problems, because it can map the network output to the probability distribution of each category. During the training process, the output of the Softmax function can be used to calculate the loss function and update the network parameters through the back propagation algorithm. It is worth noting that the output of the Softmax function does not change the relative size relationship between the elements, but only normalizes them. Therefore, the Softmax function does not change the characteristics of the input vector, but only converts it into a probability distribution form.

[0030] In the technical solution of the present application, the sequence of feature vectors describing the health care questions at word granularity and the sequence of feature vectors describing the first alternative answers at word granularity respectively express the encoded text semantic features of the health care questions and the first alternative health care answers input by the user. Thus, after the sequence of feature vectors describing the health care questions at word granularity and the sequence of feature vectors describing the first alternative answers at word granularity pass through the word granularity semantic sequence interactive fusion module, the obtained question-answer word granularity semantic interactive fusion feature vector includes not only the encoded text semantic features of the health care questions and the first alternative health care answers, but also the sequence semantic interaction features of the spatial interaction of their semantic features, that is, it has a spatial mixed feature distribution based on the semantic feature space.

[0031] However, considering the variability of the sequence of feature vectors describing the question word granularity in the health care aspect and the sequence of feature vectors describing the first alternative answer word granularity in the text semantic feature calculation dimension due to source semantic differences, it is necessary to improve the feature semantic logical dependency of the question-answer word granularity semantic interaction fusion feature vector.

[0032] In a preferred example, the question-answer word granular semantic interaction fusion feature vector is passed through a classifier to obtain a classification result, including: determining a first weight matrix and a second weight matrix, wherein the first weight matrix and the second weight matrix are The eigenvalues ​​are the first The eigenvalue and The square root of the correlation value and the L2 distance value of the eigenvalue: ;in, and They represent the first The eigenvalue and The square root of the correlation value and the L2 distance value of the eigenvalue, represents the first weight matrix Eigenvalues, represents the second weight matrix Characteristic value; respectively matrix-multiply the question-answer word granularity semantic interaction fusion feature vector with the first weight matrix to obtain a first intermediate vector, matrix-multiply the question-answer word granularity semantic interaction fusion feature vector with the second weight matrix to obtain a second intermediate vector, and dot-multiply the first weight matrix with the second weight matrix and then matrix-multiply the question-answer word granularity semantic interaction fusion feature vector to obtain a third intermediate vector; dot-multiply the first intermediate vector, the second intermediate vector and the third intermediate vector to obtain an optimized question-answer word granularity semantic interaction fusion feature vector; pass the optimized question-answer word granularity semantic interaction fusion feature vector through a classifier to obtain a classification result.

[0033] For example, the question-answer word granularity semantic interaction fusion feature vector is a row vector, then the optimized question-answer word granularity semantic interaction fusion feature vector It is expressed as: ;in, represents the question-answer word granularity semantic interaction fusion feature vector, represents the first weight matrix, represents the second weight matrix, Indicates point addition, represents matrix multiplication, represents dot product, , and represent the first intermediate vector, the second intermediate vector and the third intermediate vector respectively, Represents the optimized question-answer word granularity semantic interaction fusion feature vector.

[0034] That is, by taking the question-answer word granularity semantic interaction fusion feature vector as its statistical reference-free distribution response framework based on the self-association matrix and the self-distance matrix, and constructing a retrieval-enhanced reverse response for the question-answer word granularity semantic interaction fusion feature vector to avoid simple repetition of feature distribution, and by ensuring the intrinsic mapping logic of the question-answer word granularity semantic interaction fusion feature vector based on the retrieval-response context to avoid its surface combinatorial mapping, so as to achieve the semantic logic dependency mapping of the question-answer word granularity semantic interaction fusion feature vector while maintaining the intuitive response relevance, thereby improving the accuracy of the classification results.

[0035] Specifically, the answer returning unit 372 is used to determine whether to return the first candidate health care answer based on the classification result. That is, in a specific example, a threshold is set, and when the classification result is higher than the threshold, the first candidate answer is returned; otherwise, the first candidate answer is not returned.

[0036] It is worth mentioning that in other specific examples of the present application, it is also possible to determine whether to return the first alternative health care answer based on the question-answer word granularity semantic interaction fusion feature in other ways, for example: converting the input question into a semantic representation. Word embedding technology can be used to map each word in the question into a vector representation, and these vectors are combined in order into the input representation of the question; for the alternative health care answers, they are also converted into semantic representations; by interactively fusing the semantic representations of the question and the answer, the semantic relationship between the question and the answer is captured. An attention mechanism or other interactive model can be used to achieve semantic interactive fusion. This can enhance the semantic consistency and relevance between the question and the answer; extract features from the interactively fused semantic representation. Convolutional neural networks, recurrent neural networks, or other appropriate models can be used to extract features of questions and answers; and the similarity between questions and answers is calculated using the results of feature extraction. Cosine similarity, Euclidean distance, or other similarity measurement methods can be used to calculate the similarity score; and whether to return the first alternative health care answer is determined based on the similarity score. A threshold can be set, and when the similarity score is higher than the threshold, the first alternative answer is returned; otherwise, the first alternative answer is not returned.

[0037] It should be understood that before using the above-mentioned neural network model for inference, it is necessary to train the question semantic encoder including the word embedding layer, the answer semantic encoder including the word embedding layer, the word granularity semantic sequence interactive fusion module and the classifier. That is to say, according to the large-scale language model question-answering system 300 based on health care knowledge of the present application, it also includes a training stage 400 for training the question semantic encoder including the word embedding layer, the answer semantic encoder including the word embedding layer, the word granularity semantic sequence interactive fusion module and the classifier.

[0038] Figure 3 FIG. 1 is a block diagram of the training phase of a large-scale language model question-answering system based on health care knowledge according to an embodiment of the present application. Figure 3As shown, according to the embodiment of the present application, a large-scale language model question-answering system 300 based on health care knowledge includes: a training stage 400, including: a training data acquisition unit 410, used to obtain training data, the training data including training health care questions input by the user, and a first candidate training health care answer extracted from a background database; a training question semantic supplement unit 420, used to input the training health care question into a semantic compensator based on the ChatGPT model to obtain an optimized training health care question description; a training question semantic understanding unit 430, used to perform semantic encoding processing on the optimized training health care question description to obtain a sequence of training health care question word granularity description feature vectors; a training answer semantic understanding unit 440, used to perform semantic encoding processing on the first candidate training health care answer to obtain a sequence of first candidate training answer word granularity description feature vectors; training language The semantic interaction fusion analysis unit 450 is used to perform word granularity semantic sequence interaction fusion analysis on the sequence of word granularity description feature vectors of the training health care questions and the sequence of word granularity description feature vectors of the first alternative training answers to obtain a training question-answer word granularity semantic interaction fusion feature vector; the feature correction unit 460 is used to perform feature correction on the training question-answer word granularity semantic interaction fusion feature vector to obtain a corrected training question-answer word granularity semantic interaction fusion feature vector; the classification loss unit 470 is used to pass the corrected training question-answer word granularity semantic interaction fusion feature vector through the classifier to obtain a classification loss function value; and; the training unit 480 is used to train the question semantic encoder including the word embedding layer, the answer semantic encoder including the word embedding layer, the word granularity semantic sequence interaction fusion module and the classifier based on the classification loss function value.

[0039] Among them, the classification loss unit is used to: use the classifier to process the corrected training question-answer word granularity semantic interaction fusion feature vector to obtain a training classification result; and calculate the cross entropy loss function value between the training classification result and the true value of whether to return the first alternative health care answer as the classification loss function value.

[0040] Specifically, in a specific example, the feature correction unit is used to: fuse and correct the sequence of the training health care question word granularity description feature vectors and the sequence of the first alternative training answer word granularity description feature vectors to obtain a corrected feature vector; fuse the corrected feature vector with the training question-answer word granularity semantic interaction fusion feature vector to obtain the corrected training question-answer word granularity semantic interaction fusion feature vector.

[0041] Specifically, in a specific example, the present application preferably performs fusion correction on the sequence of the training health care question word granularity description feature vectors and the sequence of the first candidate training answer word granularity description feature vectors, which is specifically expressed as: ;in, and are the first feature vectors obtained by sequentially concatenating the feature vectors describing the problem word granularity in the training health care aspect. The second feature vector obtained by concatenating the sequence of the feature vectors describing the granularity of the first candidate training answer word The characteristic value of and denote the square of the first norm of the eigenvector and the square root of the second norm of the eigenvector respectively. The first eigenvector and the second eigenvector have the same feature vector length , represents vector addition, represents vector subtraction, and is the weight hyperparameter, represents the eigenvalue of the corrected eigenvector.

[0042] Here, the above-mentioned optimization fusion is based on the correspondence under the feature value granularity to perform a foreground manifold and a background manifold division based on the vector scale of the serialized fusion representation of the sequence of feature vectors describing the granularity of the training health care questions and the sequence of feature vectors describing the granularity of the first candidate training answer, so as to stack the dynamic feature value channelization association of the sequence of feature vectors describing the granularity of the training health care questions and the sequence of feature vectors describing the granularity of the first candidate training answer under the feature corresponding channel supermanifold aggregation mechanism, thereby marking the inter-sequence variation feature semantic information between the sequence of feature vectors describing the granularity of the training health care questions and the sequence of feature vectors describing the granularity of the first candidate training answer, and realizing a quasi-fully connected stacking fusion according to the variability of the semantic content between the sequence of feature vectors describing the granularity of the training health care questions and the sequence of feature vectors describing the granularity of the first candidate training answer in different computing dimensions. In this way, in the method composed of The correction feature vector composed of the above-mentioned correction feature vector is fused with the training question-answer word granularity semantic interaction fusion feature vector for correction, so as to improve the semantic fusion effect of the training question-answer word granularity semantic interaction fusion feature vector on the sequence of the training health care question word granularity description feature vector and the sequence of the first alternative training answer word granularity description feature vector, thereby improving the accuracy of the classification result obtained by the classifier of the question-answer word granularity semantic interaction fusion feature vector. In this way, by applying a large-scale language model question-answering system combined with health care knowledge, users can be provided with accurate and comprehensive answers and guidance to health care questions, meet the needs of users in the field of health care, thereby improving users' health awareness and quality of life, and can also provide auxiliary diagnosis and treatment suggestions for medical institutions and professionals.

[0043] As described above, the large-scale language model question and answer system 300 based on health care knowledge according to the embodiment of the present application can be implemented in various wireless terminals, such as a server with a large-scale language model question and answer algorithm based on health care knowledge. In one possible implementation, the large-scale language model question and answer system 300 based on health care knowledge according to the embodiment of the present application can be integrated into the wireless terminal as a software module and / or a hardware module. For example, the large-scale language model question and answer system 300 based on health care knowledge can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the large-scale language model question and answer system 300 based on health care knowledge can also be one of the many hardware modules of the wireless terminal.

[0044] Alternatively, in another example, the large-scale language model question and answer system 300 based on health care knowledge and the wireless terminal may also be separate devices, and the large-scale language model question and answer system 300 based on health care knowledge may be connected to the wireless terminal via a wired and / or wireless network, and transmit interactive information in accordance with an agreed data format.

[0045] The embodiments of the present disclosure have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and changes will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The selection of terms used herein is intended to best explain the principles of the embodiments, practical applications, or improvements to the technology in the market, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein.

Claims

1. A large-scale language model question-answering system based on health care knowledge, It is characterized in that include: The question collection module is used to obtain health care questions input by users; A question semantics supplement module, used for inputting the health care question into a semantic compensator based on the ChatGPT model to obtain an optimized health care question description; a question semantics understanding module, used for performing semantic encoding processing on the optimized health care question description to obtain a sequence of health care question word granularity description feature vectors; an answer acquisition module, used for extracting a first candidate health care answer from a background database; an answer semantics understanding module, used for performing semantic encoding processing on the first candidate health care answer to obtain a sequence of first candidate answer word granularity description feature vectors; a semantic interaction fusion analysis module, used for performing word granularity semantic sequence interaction fusion analysis on the sequence of health care question word granularity description feature vectors and the sequence of the first candidate answer word granularity description feature vectors to obtain question-answer word granularity semantic interaction fusion features; The answer return judgment module is used to determine whether to return the first alternative health care answer based on the question-answer word granularity semantic interaction fusion feature.

2. According to claim 1, the large-scale language model question-answering system based on health care knowledge, It is characterized in that The problem semantic understanding module is used to: perform word segmentation processing on the description of the problem in optimizing health care, and then pass it through a problem semantic encoder including a word embedding layer to obtain a sequence of feature vectors of word granularity description of the problem in health care.

3. The large-scale language model question-answering system based on health care knowledge according to claim 2, It is characterized in that The answer semantic understanding module is used to: perform word segmentation processing on the first candidate health care answer and then pass it through an answer semantic encoder including a word embedding layer to obtain a sequence of word granularity description feature vectors of the first candidate answer.

4. The large-scale language model question-answering system based on health care knowledge according to claim 3, It is characterized in that The semantic interaction fusion analysis module is used to: pass the sequence of the health care question word granularity description feature vectors and the sequence of the first alternative answer word granularity description feature vectors through a word granularity semantic sequence interaction fusion module to obtain a question-answer word granularity semantic interaction fusion feature vector as the question-answer word granularity semantic interaction fusion feature.

5. According to claim 4, the large-scale language model question-answering system based on health care knowledge, It is characterized in that The semantic interaction fusion analysis module includes: a full convolutional encoding unit for health care issues, which is used to pass the sequence of health care issue word granularity description feature vectors through a health care issue feature extractor based on a full convolutional neural network model to obtain a sequence of health care issue word granularity description full convolutional feature vectors; a first alternative answer global semantic aggregation unit, which is used to arrange the sequence of the first alternative answer word granularity description feature vectors into a first alternative answer word granularity description global feature vector; a semantic fusion unit, which is used to fuse the sequence of the health care issue word granularity description full convolutional feature vectors and the first alternative answer word granularity description global feature vector to obtain a question-answer word granularity semantic fusion feature vector; a first alternative answer global semantic association encoding unit, which is used to perform semantic association encoding on the sequence of the first alternative answer word granularity description feature vectors to obtain a first alternative answer global semantic association feature vector; a semantic feature embedding fusion unit, which is used to fuse the question-answer word granularity semantic fusion feature vector and the first alternative answer global semantic association feature vector to obtain the question-answer word granularity semantic interaction fusion feature vector.

6. The large-scale language model question-answering system based on health care knowledge according to claim 5, It is characterized in that The answer return judgment module includes: a fitness evaluation unit, which is used to pass the question-answer word granularity semantic interaction fusion feature vector through a classifier to obtain a classification result, and the classification result is used to indicate whether the fitness between the question and the answer exceeds a predetermined threshold; an answer return unit, which is used to determine whether to return the first alternative health care answer based on the classification result.

7. The large-scale language model question-answering system based on health care knowledge according to claim 6, It is characterized in that It also includes a training module for training the question semantic encoder including the word embedding layer, the answer semantic encoder including the word embedding layer, the word granularity semantic sequence interactive fusion module and the classifier.

8. The large-scale language model question-answering system based on health care knowledge according to claim 7, It is characterized in that The training module includes: a training data acquisition unit, which is used to obtain training data, wherein the training data includes training health care questions input by the user and a first candidate training health care answer extracted from a background database; a training question semantic supplement unit, which is used to input the training health care questions into a semantic compensator based on a ChatGPT model to obtain an optimized training health care question description; a training question semantic understanding unit, which is used to perform semantic encoding processing on the optimized training health care question description to obtain a sequence of training health care question word granularity description feature vectors; a training answer semantic understanding unit, which is used to perform semantic encoding processing on the first candidate training health care answer to obtain a sequence of first candidate training answer word granularity description feature vectors; a training semantic interaction fusion analysis unit, which is used to perform semantic encoding processing on the training health care question description to obtain a sequence of first candidate training answer word granularity description feature vectors; A sequence of word granularity description feature vectors of the aspect question and a sequence of word granularity description feature vectors of the first alternative training answer are subjected to word granularity semantic sequence interaction fusion analysis to obtain a training question-answer word granularity semantic interaction fusion feature vector; a feature correction unit is used to perform feature correction on the training question-answer word granularity semantic interaction fusion feature vector to obtain a corrected training question-answer word granularity semantic interaction fusion feature vector; a classification loss unit is used to pass the corrected training question-answer word granularity semantic interaction fusion feature vector through the classifier to obtain a classification loss function value; and a training unit is used to train the question semantic encoder including a word embedding layer, the answer semantic encoder including a word embedding layer, the word granularity semantic sequence interaction fusion module and the classifier based on the classification loss function value.

9. The large-scale language model question-answering system based on health care knowledge according to claim 8, It is characterized in that The classification loss unit is used to: use the classifier to process the corrected training question-answer word granularity semantic interaction fusion feature vector to obtain a training classification result; and calculate the cross entropy loss function value between the training classification result and the true value of whether to return the first alternative health care answer as the classification loss function value.

Images