Automatic construction system of medical intelligent knowledge graph based on end-to-end differential learning

Abstract

The present invention provides an automatic construction system of medical intelligent knowledge graph based on end-to-end differential learning, including an end-to-end differential learning framework and a hierarchical multi-task entity recognition module, a graph-enhanced relationship extraction module, and a large model-assisted verification and optimization module integrated in the end-to-end differential learning framework; the hierarchical multi-task entity recognition module, the graph-enhanced relationship extraction module, and the large model-assisted verification and optimization module are sequentially differentiably connected. The present invention can form a closed-loop knowledge extraction-verification-optimization process, and realize the full-process differentiable automated and precise conversion from unstructured medical text to high-quality knowledge graph.

Description

Technical Field

[0001] The present invention relates to the technical field of text processing, and in particular to an automatic construction system of a medical intelligent knowledge graph based on end-to-end differential learning. Background Art

[0002] As a structured knowledge representation method, knowledge graph plays an important role in professional fields such as medicine. At present, the construction of medical knowledge graph mainly adopts a pipeline processing solution, which usually includes multiple processing modules such as entity recognition, relationship extraction and knowledge verification. In this solution, each module is trained and optimized independently, and each module is responsible for a specific task, such as the entity recognition module identifies medical entities in the text, the relationship extraction module analyzes the semantic relationship between entities, and the knowledge verification module verifies the extracted knowledge. Although this modular processing method is relatively simple in engineering implementation, it has serious technical limitations.

[0003] The most prominent problem is the separation between modules. In the prior art, there is a lack of effective information exchange mechanism between the various processing modules, and the errors generated by the upstream modules will continue to accumulate and amplify in the downstream modules. For example, when errors occur in the entity recognition module, these errors will directly affect the performance of the relationship extraction module, and then affect the accuracy of knowledge verification. More seriously, due to the separation between modules, the feedback information obtained by the downstream module cannot be effectively used to optimize the performance of the upstream module.

[0004] Another key issue is the discrete decision-making characteristics in the process of building the knowledge graph. In traditional solutions, the determination of entity boundaries, the determination of entity types, the selection of relationship types, etc. are all discrete decision-making processes. This discreteness makes it difficult for the entire system to achieve end-to-end gradient backpropagation, and thus cannot be jointly optimized. At the same time, knowledge verification is usually a post-processing step, completely separated from the knowledge extraction process. The verification results cannot be fed back to the knowledge extraction stage, and there is also a lack of a mechanism for real-time verification during the extraction process.

[0005] In addition, each module in the prior art has its own independent optimization goal. The entity recognition module is committed to improving the accuracy of entity recognition, the relationship extraction module focuses on the accuracy of relationship classification, and the knowledge verification module focuses on the accuracy of verification. This split optimization goal leads to local optimization that may damage the overall effect, and lacks a unified, end-to-end optimization goal.

[0006] The above problems have seriously affected the quality and efficiency of medical knowledge graph construction. At present, there is a lack of a unified solution that can achieve end-to-end differential learning and promote the coordinated optimization of various modules. This requires an innovative technical solution that can achieve differentiable optimization of the entire process, break through the information barriers between modules, support end-to-end gradient backpropagation, and build an efficient inter-module coordination mechanism. This solution should not only be able to solve the module splitting problem in the existing technology, but also be able to effectively deal with the optimization difficulties brought about by discrete decisions, and achieve overall optimization of the knowledge graph construction process. Summary of the invention

[0007] The present invention aims to solve at least one of the technical problems in the above-mentioned technologies to a certain extent. To this end, the purpose of the present invention is to propose an automatic construction system of medical intelligent knowledge graph based on end-to-end differential learning, which can automatically convert unstructured medical text into high-quality medical knowledge graph in a full-process differentiable form.

[0008] To achieve the above-mentioned purpose, an embodiment of the present invention proposes a medical intelligent knowledge graph automatic construction system based on end-to-end differential learning, including an end-to-end differential learning framework and a hierarchical multi-task entity recognition module, a graph-enhanced relationship extraction module, and a large model-assisted verification and optimization module integrated in the end-to-end differential learning framework; the hierarchical multi-task entity recognition module, the graph-enhanced relationship extraction module, and the large model-assisted verification and optimization module are sequentially differentiably connected;

[0009] The hierarchical multi-task entity recognition module includes a BioBERT encoder, a bidirectional GRU layer, an attention mechanism, a hierarchical classifier, and a conditional random field layer connected in sequence, and is used to capture deep semantic features, capture long-term dependency relationships, dynamically assign different weights to different parts of the input sequence according to the context, perform hierarchical classification, and calculate the optimal label combination of the sequence according to the input pre-processed medical text sequence, so as to identify and extract various medical entities;

[0010] The graph-enhanced relationship extraction module includes a context encoder, a graph structure, a graph attention network, and a multi-task classifier connected in sequence, which are used to capture contextual semantic information processing, construct a sentence-level graph structure processing, capture semantic relationship processing at different levels through a multi-layer graph attention mechanism, and perform relationship classification and entity type prediction processing according to the input medical entity, so as to identify and extract the semantic relationship between medical entities and output relationship-enhanced text;

[0011] The large model-assisted verification and optimization module is used to verify and optimize the relationship-enhanced text representation using a large-scale speech model.

[0012] According to the automatic construction system of medical intelligent knowledge graph based on end-to-end differential learning of the embodiment of the present invention, by integrating three core modules of hierarchical multi-task entity recognition module, graph-enhanced relationship extraction module and large model-assisted verification and optimization module in the end-to-end differential learning framework, it can form a closed-loop knowledge extraction-verification-optimization process, thereby realizing the full-process differentiable and automated conversion from unstructured medical text to high-quality knowledge graph, and achieving significant improvements in entity recognition accuracy, relationship extraction completeness and knowledge graph quality. The automatic construction system of medical intelligent knowledge graph based on end-to-end differential learning of the embodiment of the present invention not only improves the efficiency and quality of medical knowledge graph construction, but also can provide valuable reference for cross-domain knowledge extraction and graph construction tasks.

[0013] In addition, the medical intelligent knowledge graph automatic construction system based on end-to-end differential learning proposed in the above embodiment of the present invention may also have the following additional technical features:

[0014] Optionally, the BioBERT encoder comprises:

[0015] The first multi-layer Transformer encoder, each layer of which has a self-attention mechanism and a feed-forward neural network; and

[0016] The context-sensitive representation unit is used to capture long-distance dependencies in the sequence through the self-attention mechanism and generate context-sensitive word representations.

[0017] Optionally, the bidirectional GRU layer includes:

[0018] A bidirectional processing unit for processing input sequences bidirectionally and in parallel, capturing long-range dependencies in the sequences and generating context-sensitive word representations; and

[0019] The gate control mechanism unit includes an update gate and a reset gate, and is used to dynamically control the flow of information during the processing of the bidirectional processing unit.

[0020] Optionally, the attention mechanism comprises:

[0021] A multi-angle feature extraction unit, which uses multiple attention heads to extract different features of the input sequence at the same time;

[0022] The attention weight calculation unit is used to dynamically calculate the weights of different features according to the context through multi-head splicing and linear transformation.

[0023] Optionally, the conditional random field layer includes:

[0024] A transition probability matrix unit, used to learn the transition probability between labels and obtain a transition matrix; and

[0025] The emission probability unit is used to combine the output of the hierarchical classifier and the transfer matrix to calculate the optimal label sequence.

[0026] Optionally, the context encoder comprises

[0027] The second multi-layer Transformer encoder is used to generate context-sensitive word representations;

[0028] An attention pooling unit is used to pool the output of the second multi-layer Transformer encoder using an attention mechanism to obtain a sentence-level representation.

[0029] Optionally, the graph attention network comprises a plurality of graph attention layers;

[0030] Each of the multiple graph attention layers adopts residual connection, and each graph attention layer adopts a multi-head attention mechanism; wherein the attention calculation process of the multi-head attention mechanism will consider the type of edge.

[0031] Optionally, the multi-task classifier adopts a multi-instance learning method based on an attention mechanism to calculate, for each entity pair, the attention weights of all sentences containing the entity pair, and perform a final relationship prediction based on the weighted instance representation.

[0032] Optionally, a preprocessing module is further included; the preprocessing module is integrated into the end-to-end differential learning framework and connected to the input end of the hierarchical multi-task entity recognition module;

[0033] The preprocessing module is used to convert the original unstructured medical text into a standardized, structured preprocessed medical text sequence.

[0034] Optionally, the preprocessing module includes a text cleaning unit, a sentence segmentation unit, a word standardization unit, a part-of-speech tagging unit and a dependency syntax analysis unit. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 A simplified schematic diagram of the architecture of a medical intelligent knowledge graph automatic construction system based on end-to-end differential learning provided by an embodiment of the present invention;

[0036] Figure 2 A schematic diagram of the complete architecture of a system for automatically constructing a medical intelligent knowledge graph based on end-to-end differential learning provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0037] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and should not be construed as limiting the present invention.

[0038] In order to better understand the above technical solution, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

[0039] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0040] Figure 1 A simplified architecture diagram of a system for automatically constructing a medical intelligent knowledge graph based on end-to-end differential learning provided in an embodiment of the present invention.

[0041] like Figure 1 As shown, an embodiment of the present invention provides an Intelligent Medical Knowledge Auto-Construction System based on End-to-End Differentiable Learning (IMKACS-E2DL for short), which includes an end-to-end differential learning framework (E2DL), a hierarchical multi-task entity recognition module (HMT-NER), a graph enhanced relationship extraction module (GARE) and a large model assisted verification and optimization module (LLMVO); the hierarchical multi-task entity recognition module (HMT-NER), the graph enhanced relationship extraction module (GARE) and the large model assisted verification and optimization module (LLMVO) are sequentially differentiably connected and integrated in the end-to-end differential learning framework (E2DL).

[0042] The hierarchical multi-task entity recognition module (HMT-NER) mainly includes a BioBERT encoder, a bidirectional GRU layer, an attention mechanism, a hierarchical classifier and a conditional random field layer connected in sequence, which is used to capture deep semantic features, capture long-term dependency relationships, dynamically assign different weights to different parts of the input sequence according to the context, perform hierarchical classification and calculate the optimal label combination of the sequence according to the input preprocessed medical text sequence, so as to identify and extract various medical entities.

[0043] The graph enhanced relationship extraction module (GARE) includes a context encoder, a graph structure, a graph attention network, and a multi-task classifier connected in sequence, which are used to capture contextual semantic information processing, construct sentence-level graph structure processing, capture semantic relationship processing at different levels through a multi-layer graph attention mechanism, and perform relationship classification and entity type prediction processing according to the input medical entity, so as to identify and extract the semantic relationship between medical entities and output relationship enhanced text.

[0044] The large model-assisted verification and optimization module (LLMVO) is used to verify and optimize the relation-enhanced text representation using a large-scale speech model.

[0045] In this embodiment, the end-to-end differential learning framework (E2DL) can convert discrete module decisions into continuous soft decisions, and realize the full differentiable optimization of the entire medical knowledge graph construction process. Specifically, the framework significantly enhances the synergy between the modules integrated therein through a unified loss function and innovative optimization strategies (such as soft attention entity alignment, relationship-aware entity tuning, large model-guided noise contrast learning, etc.), and improves the overall performance and robustness of the system.

[0046] The Hierarchical Multi-Task Entity Recognition (HMT-NER) module combines BioBERT, bidirectional GRU and conditional random fields, and adopts a hierarchical classification strategy to simultaneously predict the main type and subtype of an entity. In particular, the innovative introduction of attention mechanism and dynamic context window improves the recognition accuracy of complex medical entities.

[0047] The graph-enhanced relation extraction module (GARE) constructs an innovative hierarchical graph structure based on the graph attention network, and combines the multi-head self-attention mechanism to effectively capture multi-hop relationships and global dependencies between entities. In particular, the dynamic graph structure learning and multi-granularity graph convolution operations introduced in the hierarchical graph structure improve the accuracy and completeness of relation extraction.

[0048] The large model-assisted verification and optimization module (LLMVO) innovatively uses large-scale language models for knowledge verification and reasoning; through differentiable verification scoring, adversarial verification and knowledge distillation technology, it not only improves the reliability of extracted knowledge, but also discovers potential new knowledge and realizes automatic expansion and optimization of the knowledge graph.

[0049] The medical intelligent knowledge graph automatic construction system based on end-to-end differential learning provided in this embodiment performs unified optimization of the hierarchical multi-task entity recognition module (HMT-NER), graph enhanced relationship extraction module (GARE) and large model assisted verification and optimization module (LLMVO) integrated therein through the end-to-end differential learning framework (E2DL), which can give full play to the advantages of each module and enhance the synergy between modules. Experiments show that compared with traditional methods and independently optimized modular systems, this embodiment can achieve significant improvements in entity recognition accuracy, relationship extraction completeness and knowledge graph quality. Therefore, this embodiment not only improves the efficiency and quality of medical knowledge graph construction, but also provides a high-value reference for cross-domain knowledge extraction and graph construction tasks.

[0050] See also Figure 2 , Figure 2 A schematic diagram of the complete architecture of a medical intelligent knowledge graph automatic construction system based on end-to-end differential learning provided by an embodiment of the present invention. Figure 1 The embodiment is further expanded to provide a system for automatically constructing a medical intelligent knowledge graph based on end-to-end differential learning.

[0051] like Figure 2 As shown, in this embodiment, the medical intelligent knowledge graph automatic construction system based on end-to-end differential learning also includes: a preprocessing module. The preprocessing module is also integrated into the end-to-end differential learning framework and connected to the input end of the hierarchical multi-task entity recognition module.

[0052] The preprocessing module is used to convert the original unstructured medical text into a standardized, structured preprocessed medical text sequence, laying the foundation for subsequent entity recognition and relationship extraction tasks.

[0053] In some specific embodiments, the preprocessing module specifically includes a text cleaning unit, a sentence segmentation unit, a word standardization unit, a part-of-speech tagging unit, and a dependency syntactic analysis unit to respectively implement text cleaning, sentence segmentation, word standardization, part-of-speech tagging, and dependency syntactic analysis processing. In particular, the preprocessing module optimizes each unit therein to better adapt to the characteristics of medical texts.

[0054] In the above specific implementation, the text cleaning unit is used to eliminate noise and non-standard expressions in the original medical text that may affect subsequent processing.

[0055] Specifically, the text cleaning unit is configured to perform the following operations:

[0056] a) Remove extra whitespace characters and special symbols: Use regular expressions to match and remove unnecessary spaces, tabs, newlines, etc. in the text. For example, convert "Hypertension is a common disease." to "Hypertension is a common disease.".

[0057] b) Unify punctuation marks: Convert full-width symbols to half-width symbols to ensure the consistency of punctuation marks. For example, convert "hypertension," to "hypertension,"

[0058] c) Correct common spelling errors and irregular abbreviations: Use a pre-prepared dictionary of medical terms to automatically correct common spelling errors, for example, automatically correct "diebetes" to "diabetes".

[0059] The text cleaning unit relies on a powerful rule engine and a comprehensive medical terminology dictionary to achieve the above operations. The rule engine contains a series of regular expressions and transformation rules, while the medical terminology dictionary contains a large number of standard medical terms and their common misspelling variants. Therefore, through the text cleaning unit, the special properties of medical texts, such as professional terms, abbreviations, complex sentence structures, etc., can be effectively processed, thereby improving the performance and accuracy of the entire system.

[0060] In the above specific implementation, the sentence segmentation unit is used to segment the medical text using a rule-based and machine learning method. Medical texts often contain complex sentence structures, and accurate sentence segmentation is crucial for subsequent processing.

[0061] Specifically, the sentence unit is configured to be able to perform the following operations:

[0062] a) Rule-based preliminary sentence segmentation: Use regular expressions to identify common sentence boundary markers, such as period, question mark, exclamation mark, etc. But be careful to handle special cases, such as the period in an abbreviation (such as "et al.") should not be regarded as the end of a sentence.

[0063] b) Machine learning model optimization: A BiLSTM-CRF-based sequence labeling model is used to optimize the initial sentence segmentation results. This model has been trained with a large amount of medical text and can accurately identify sentence boundaries in complex situations.

[0064] For example, for the text "The patient presented with severe abdominal pain. Dr. Smith diagnosed acute appendicitis. The treatment plan includes immediate surgery.", our sentence segmentation module can correctly identify three sentences and correctly handle abbreviations such as "Dr."

[0065] In the above specific implementation, the word standardization unit is used to unify the abbreviations and synonyms in the medical text to improve the consistency of subsequent processing. There are a large number of abbreviations and synonyms in the medical field, and the word standardization unit can unify these expressions.

[0066] Specifically, the word standardization unit is configured to perform the following operations:

[0067] a) Abbreviation expansion: Build a comprehensive dictionary of medical abbreviations and expand abbreviations in the text to their full names, for example, expanding "HTN" to "hypertension".

[0068] b) Synonym unification: Use a synonym dictionary to unify different expressions of the same concept into a standard form. For example, unify "heart attack" and "myocardial infarction" into "myocardial infarction".

[0069] The word standardization unit implements the above steps by constructing a high-quality abbreviation dictionary and a synonym dictionary. Preferably, a semi-automatic method is used to construct the above dictionary: first, initial entries are extracted from authoritative medical resources (such as UMLS), then new entries are automatically extracted from large-scale medical literature through a machine learning model, and finally reviewed and supplemented by medical experts.

[0070] In the above specific implementation, the part-of-speech tagging unit and the dependency syntax analysis unit are used to provide richer linguistic features for subsequent entity recognition and relationship extraction tasks.

[0071] Specifically, the part-of-speech tagging unit is configured to perform the following operations:

[0072] Part-of-speech tagging: Use a sequence tagging model based on BiLSTM-CRF to tag each word. This model has been trained with a large amount of medical text and can accurately identify special parts of speech such as medical proper nouns and drug names.

[0073] Specifically, the dependency syntax analysis unit is configured to perform the following operations:

[0074] Dependency parsing: A neural network-based dependency parser is used to construct a grammatical structure tree for the sentence. The dependency parser is optimized to handle long sentences and complex clause structures commonly found in medical texts.

[0075] For example, for the sentence "Aspirin is commonly used to treat fever and mild to moderate pain.", our system can correctly tag "Aspirin" as a proper noun (NNP), "treat" as a verb (VB), and recognize that "Aspirin" is the subject of "used" and "treat" is the purpose complement of "used".

[0076] In this embodiment, a more accurate preprocessed text representation (PTR) is obtained through the preprocessing module:

[0077] PTR = {

[0078] 'original_text': str,

[0079] 'cleaned_text': str,

[0080] 'sentences': List[Dict{

[0081] 'text': str,

[0082] 'standardized_words': List[str],

[0083] 'pos_tags': List[str],

[0084] 'dependency_tree': List[Tuple[int, int, str]]

[0085] }]

[0086] }

[0087] in,

[0088] 'original_text' represents the original input text string;

[0089] 'cleaned_text' represents the cleaned text string;

[0090] 'sentences' represents a list of sentences, each sentence is a dictionary;

[0091] Each sentence dictionary contains:

[0092] 'text': the text string of the sentence;

[0093] 'standardized_words': List of words that have been standardized;

[0094] 'pos_tags': list of part-of-speech tags;

[0095] 'dependency_tree': Dependency parsing results, represented by a list of tuples of (child node index, parent node index, relationship type).

[0096] In this embodiment, the Hierarchical Multi-Task Named Entity Recognition (HMT-NER), namely Hierarchical Multi-Task Named Entity Recognition, HMT-NER, referred to as GARE, is the primary core component of the embodiment of the present invention. Its main goal is to accurately identify and extract various medical entities, such as diseases, symptoms, drugs, genes, etc., from the preprocessed medical text (i.e., the output of the above-mentioned preprocessing module).

[0097] Here, the hierarchical multi-task entity recognition module (HMT-NER) adopts advanced deep learning technology and combines it with medical field-specific optimization strategies to address the special challenges of entity recognition in medical texts, such as the complexity of professional terms, the existence of nested entities, and strong contextual dependencies. Through model architecture design and training strategies, the hierarchical multi-task entity recognition module (HMT-NER) can efficiently and accurately recognize various medical entities.

[0098] In this embodiment, the hierarchical multi-task entity recognition module (HMT-NER) is based on the deep learning model of BioBERT, which is specifically optimized and extended to adapt to the characteristics of medical texts. Specifically, the hierarchical multi-task entity recognition module (HMT-NER) includes: input representation, BioBERT encoder, bidirectional GRU layer, attention mechanism, hierarchical classifier, conditional random field (CRF) layer, model training and optimization, nested entity processing, and post-processing optimization.

[0099] like Figure 2 As shown, the overall architecture of the BioBERT deep learning model can be expressed as:

[0100] Input (input representation) -> BioBERT (BioBERT encoder) -> BiGRU (bidirectional GRU layer) -> Attention (attention mechanism) -> Hierarchical Classifier (hierarchical classifier) ​​-> CRF Layer (conditional random field (CRF) layer) -> Output.

[0101] The input representation specifically adopts the following strategies to enhance the input identification capability:

[0102] a) WordPiece segmentation: Use BioBERT's WordPiece segmenter to segment the input text into subword units. This method can effectively handle rare words and compound words in the medical field. For example, "hypertension" may be segmented into "hyper" and "##tension".

[0103] b) Position encoding: Add position information to each word to help the model understand the relative position of the word in the sentence.

[0104] c) Part-of-speech tag embedding: The part-of-speech tag information obtained in the preprocessing stage is encoded into a vector and added to the word representation. This provides additional grammatical information for the model.

[0105] As a specific example, the input representation can be formally expressed as:

[0106]

[0107] in, is the input word sequence, is the corresponding part-of-speech tag sequence, is the final input representation.

[0108] In this embodiment, the BioBERT encoder is used to capture the deep semantic features of the input sequence. It is different from the common BERT in the prior art. It is a version of the BERT model that continues to be pre-trained on large-scale biomedical literature and has a better understanding of medical terms and expressions.

[0109] The BioBERT encoder comprises:

[0110] a) The first multi-layer Transformer encoder, each layer of which has a self-attention mechanism and a feed-forward neural network;

[0111] b) Context-sensitive representation unit, which is used to capture long-distance dependencies in the sequence through the self-attention mechanism and generate context-sensitive word representations.

[0112] As a specific example, the encoding process of the BioBERT encoder can be expressed as:

[0113]

[0114] in, It is Context-sensitive representation of words.

[0115] In this embodiment, the bidirectional GRU layer, also known as the bidirectional gated recurrent unit (BiGRU) layer, is specifically located above the output of the BioBERT encoder. The bidirectional GRU layer is a variant of the recurrent neural network and has the ability to capture long-term dependencies.

[0116] In particular, by adding a new bidirectional GRU layer after the BioBERT encoding, the model's ability to capture long-distance dependencies and sequence information in sequences can be further enhanced, which helps to recognize long entity names and is very helpful for processing complex medical terminology.

[0117] The bidirectional GRU layer mainly includes:

[0118] a) A bidirectional processing unit, which is used to process the input sequence in parallel in both directions, capture the long-distance dependencies in the sequence, and generate context-sensitive word representations. Here, the bidirectional parallelism specifically refers to processing from left to right and from right to left at the same time. This captures more comprehensive context information, which is particularly important for understanding the contextual relationship of medical terms.

[0119] b) A gating mechanism unit, including an update gate and a reset gate, for dynamically controlling the flow of information during the processing of the bidirectional processing unit. Here, the update gate can determine how much previous information to retain; the reset gate can determine how much previous information to ignore. This mechanism enables the model to better handle long-term dependencies in long sequences.

[0120] The bidirectional GRU layer has a parameter efficiency comparable to that of the existing LSTM, but has the advantages of fewer parameters and faster training speed.

[0121] As a specific example, the mathematical expression of the bidirectional GRU layer is as follows:

[0122] For the forward GRU:

[0123] in, Indicates time The input vector of Indicates time The hidden state vector of Represents the update gate vector, which controls the degree of retention of the hidden state at the previous moment; Represents the reset gate vector, which controls the influence of the previous hidden state on the current candidate hidden state; represents the candidate hidden state vector; They are the weight matrices of the update gate, reset gate and candidate hidden state respectively.

[0124] For the backward GRU, the computation is similar, but starts from the end of the sequence.

[0125] The final BiGRU output is the concatenation of the forward and backward GRU outputs:

[0126]

[0127] in, is the output of BioBERT, is the output of BiGRU, It is The representation of a word after BiGRU processing contains forward and backward information.

[0128] In this embodiment, the introduction of the bidirectional GRU layer can significantly improve the model's capabilities in the following aspects:

[0129] (1) Long-distance dependency modeling: For long entity names (such as certain rare diseases or complex biomarkers), the bidirectional GRU layer can better capture their internal structure and contextual relationships.

[0130] (2) Sequence labeling consistency: Through bidirectional processing, the bidirectional GRU layer helps maintain labeling consistency and reduce unreasonable labeling sequences such as "B-Disease I-Symptom".

[0131] (3) Context-sensitive representation: The representation generated by the bidirectional GRU layer integrates the information of the entire sequence and helps to resolve ambiguous problems, such as distinguishing whether "psoriasis" refers to a disease or a symptom.

[0132] In this embodiment, by introducing an attention mechanism into the hierarchical multi-task entity recognition module (HMT-NER), the model can dynamically focus on important parts of the input sequence. The attention mechanism enables the model to dynamically assign different weights to different features of the input sequence according to the context in different prediction tasks, thereby better focusing on key information related to entity recognition.

[0133] The attention mechanism described is a multi-head self-attention mechanism, which mainly includes:

[0134] a) A multi-angle feature extraction unit, which is used to extract different features of the input sequence simultaneously using multiple attention heads, so that the model can focus on different aspects of the sequence at the same time, such as grammatical structure, semantic relationship, etc.

[0135] b) Attention weight calculation unit, used to dynamically calculate the weights of different features according to the context through multi-head splicing and linear transformation. Specifically, the attention weight calculation unit has long-distance dependency modeling and parallel computing capabilities. The former can directly establish the connection between any two positions and effectively capture long-distance dependencies; the latter can improve computing efficiency compared to traditional recurrent neural networks.

[0136] As a specific example, the mathematical expression of the attention mechanism is as follows:

[0137] For each attention head , in is the input sequence (the concatenated output of BioBERT and BiGRU), is a learnable weight matrix.

[0138] Calculate the attention weights: in is the dimension of the key vector. Compute the weighted sum: .

[0139] Multi-head splicing and linear transformation: ,in is the weight matrix of the output linear transformation.

[0140] The final attention output can be simply expressed as:

[0141]

[0142] in, is the original input (the concatenated output of BioBERT and BiGRU), It is the feature representation after attention mechanism processing.

[0143] In this specific implementation, the introduction of the attention mechanism can bring the following advantages:

[0144] (1) Dynamic context understanding: For a word like “liver”, the attention mechanism can dynamically decide whether to recognize it as an organ name or part of “hepatitis” based on the context.

[0145] (2) Integration of distant information: When processing long texts, the attention mechanism can integrate distant related information, such as linking the symptoms mentioned at the beginning of the article with the diagnosis later in the text.

[0146] (3) Accurately locate entity boundaries: By focusing on the start and end tags of an entity, the attention mechanism helps to locate entity boundaries more accurately, especially for the recognition of nested entities.

[0147] (4) Improve model interpretability: By analyzing the attention weights, we can understand which parts of the input the model focuses on when making predictions, thus improving the interpretability of the model.

[0148] In this embodiment, the hierarchical classifier is a hierarchical classifier specially designed in consideration of the complex hierarchical structure of medical entities. The hierarchical classifier is used to first predict the main type of entity (such as disease, symptom, drug, etc.), and then predict a more fine-grained subtype based on the main type.

[0149] As a specific example, the mathematical expression of the hierarchical classifier is as follows:

[0150]

[0151] in, Represents a subtype. is the prediction result of the main type, It is the prediction result of sub-types under each main type.

[0152] In this embodiment, the conditional random field layer is also called a CRF layer. The CRF layer can receive the output of the above-mentioned hierarchical classifier as input, and effectively improve the accuracy of sequence labeling by modeling the joint probability distribution of the label sequence; especially when processing the boundary recognition of medical entities, it can effectively improve the accuracy of medical entity boundary recognition.

[0153] The conditional random field layer mainly includes:

[0154] a) The transition probability matrix unit is used to learn the transition probability between labels, such as the B-Disease label is more likely to be followed by the I-Disease label, and obtain the transition matrix;

[0155] b) A transmission probability unit, used to combine the output of the hierarchical classifier and the above transfer matrix to calculate the optimal label sequence.

[0156] As a specific example, the conditional random field layer uses the following formula to calculate the label sequence Given input The conditional probability of:

[0157]

[0158] in, is a normalization factor equal to the sum of the scores of all possible label sequences.

[0159] The conditional random field layer uses the output of the hierarchical classifier and its own transfer matrix to calculate the most likely label sequence. Its output is the optimal label combination of the entire sequence.

[0160] In this embodiment, the model training and optimization refers to training and entity recognition of the hierarchical multi-task entity recognition module.

[0161] The hierarchical multi-task entity recognition module is trained using a medical text dataset with entity annotations. The training process adopts a multi-task learning framework, and the loss function includes main type classification loss, subtype classification loss and CRF loss:

[0162] in, is the cross entropy loss for main type classification, is the cross entropy loss for subtype classification, is the negative log-likelihood loss of the CRF layer, and is the weight coefficient for balancing different tasks.

[0163] In the inference phase, the model processes the input text and outputs the entity label of each word (such as B-Disease, I-Disease, O, etc.).

[0164] The output of the hierarchical multi-task entity recognition module will be formatted into a new data structure called "Entity-Enhanced Text Representation" (EETR):

[0165] As a specific example, EETR = {

[0166] 'original_text': str,

[0167] 'cleaned_text': str,

[0168] 'sentences': List[Dict{

[0169] 'text': str,

[0170] 'tokens': List[str],

[0171] 'pos_tags': List[str],

[0172] 'dependency_tree': Dict{int: int},

[0173] 'entities': List[Dict{

[0174] 'text': str,

[0175] 'start': int,

[0176] 'end': int,

[0177] 'main_type': str,

[0178] 'sub_type': str,

[0179] 'confidence': float

[0180] }]

[0181] }]

[0182] }

[0183] This EETR structure adds an 'entities' field to each sentence based on the original preprocessed text representation (PTR), which contains the recognized entity information, including the main type, subtype and confidence.

[0184] In this embodiment, the hierarchical multi-task entity recognition module (HMT-NER) adopts an innovative "Hierarchical Multi-Task Medical Named Entity Recognition Model" (HMT-MNER) architecture, which combines the powerful feature extraction capability of the pre-trained language model, the sequence modeling capability of the recurrent neural network, the key information capture capability of the attention mechanism, and the sequence labeling advantage of the conditional random field.

[0185] In this embodiment, the graph-augmented relation extraction module (GARE), namely Graph-Augmented Relation Extraction, is referred to as GARE. As another core component of the automatic construction system of the medical intelligent knowledge graph based on end-to-end differential learning, it is mainly used to identify and extract semantic relationships between identified entities. In the medical field, these relationships may include treatment relationships (such as drug treatment of diseases), etiology relationships (such as genes causing diseases), symptom relationships (such as symptoms manifesting diseases), etc. The graph-augmented relation extraction module (GARE) adopts deep learning technology, combined with graph neural networks and attention mechanisms, which can effectively capture complex semantic relationships and long-distance dependencies to accurately identify complex relationships between various medical entities.

[0186] The core of the Graph Enhanced Relation Extraction Module (GARE) of this embodiment is a deep learning model based on the Graph Attention Network (GAT), which has been specifically optimized and extended to adapt to the complex relational expressions in medical texts.

[0187] The main components of the graph-enhanced relation extraction module (GARE) include: input representation, context encoder, graph construction, graph attention network, relation classification, and multi-task learning framework.

[0188] like Figure 2 As shown, the overall structure of the model in the graph enhanced relation extraction module (GARE) can be expressed as:

[0189] Input -> BioBERT (Context Encoder) -> Graph Construction -> Hierarchical Graph Attention Network HGAT -> Multi-task Classifier -> Output.

[0190] In this embodiment, the input representation refers to receiving the EETR (Entity Enhanced Text Representation) output by the Hierarchical Multi-Task Entity Recognition Module (HMT-NER) as input.

[0191] In some embodiments, the graph enhanced relation extraction module (GARE) enhances the representation capability of its input through the following strategies:

[0192] a) Word embedding: Use BioBERT’s word embedding layer to obtain basic word representations.

[0193] b) Position embedding: Add position information relative to the target entity pair to each word to help the model understand the relative position of the word and the entity.

[0194] c) Entity type embedding: The predicted type of the entity (such as disease, drug, gene, etc.) is encoded as a vector and added to the word representation at the corresponding position.

[0195] As a specific example, the input representation can be formally expressed as:

[0196]

[0197] in, is the input word sequence, and is the target entity pair, is a syntactic dependency tree, is the final input representation.

[0198] In this embodiment, the context encoder uses the BioBERT encoder as a basic encoder to capture the deep semantic features and long-distance dependencies of the input text sequence, that is, contextual semantic information.

[0199] The context encoder specifically comprises:

[0200] a) A second multi-layer Transformer encoder, which uses the multi-layer Transformer structure of the BioBERT encoder to generate context-sensitive word representations;

[0201] b) An attention pooling unit, which is used to pool the output of the second multi-layer Transformer encoder using an attention mechanism to obtain a sentence-level representation.

[0202] As a specific example, the context encoding process of the context encoder can be expressed as:

[0203]

[0204] in, is the word-level context representation, It is a sentence-level representation.

[0205] The specific calculation of attention pooling is as follows:

[0206]

[0207] in, and is the learnable parameter matrix.

[0208] In this embodiment, the graph structure, also called graph construction, is a specially constructed sentence-based graph structure that can better capture the interactions and global dependencies between entities.

[0209] During the construction of the graph structure:

[0210] a) Node definition: The nodes of the graph include all the words in the sentence and special sentence nodes. The initial feature of each node is its corresponding BioBERT output.

[0211] b) Edge definition: Define multiple types of edges to capture relationships at different levels:

[0212] Grammatical dependency edges: constructed based on the results of dependency syntactic analysis.

[0213] Adjacent word edges: connect adjacent words and capture local context.

[0214] Intra-entity edges: connect words belonging to the same entity.

[0215] Candidate relationship edges: connect pairs of entities that may have a relationship.

[0216] Global edges: All nodes are connected to special sentence nodes for global propagation of information.

[0217] c) Edge weight initialization: Assign an initial weight to each edge based on the type of edge and the information of the connected nodes.

[0218] For example, the weight of a grammatical dependency edge might be based on the type of the dependency relation, and the weight of a candidate relation edge might be based on the type combination of the entity pair.

[0219] As a specific example, the mathematical formula of the graph construction process can be expressed as:

[0220]

[0221] in, is the dependency syntactic tree of the sentence, is the set of recognized entities.

[0222] In this embodiment, the graph attention network, also known as the hierarchical graph attention network (HGAT), is the core innovation of the graph enhanced relationship extraction module, which can gradually capture semantic relationships at different levels through a multi-layer graph attention mechanism:

[0223] a) Multi-layer structure: The graph attention network contains multiple graph attention layers, and each layer further refines the node representation based on the previous layer.

[0224] b) Multi-head attention: Each graph attention layer adopts a multi-head attention mechanism to learn the relationship between nodes from different representation subspaces.

[0225] c) Edge type awareness: The type of edge is considered during the attention calculation process, allowing the model to distinguish different types of relations.

[0226] d) Residual Connection: Each layer uses residual connection, which helps to train deeper networks and preserve low-level features.

[0227] As a specific example, the graph attention network is The graph attention of the layer is calculated as:

[0228]

[0229] in, Is a node Features, It is the edge The type of embedding, and is a learnable weight matrix, is the attention vector.

[0230] The final node representation is the concatenation of all layer outputs:

[0231]

[0232] The above-mentioned graph attention network uses GAT to learn the interactions and global dependencies between entities based on the sentence-level graph structure.

[0233] In this embodiment, the multi-task classifier adopts a multi-instance learning method based on an attention mechanism to calculate the attention weights of all sentences containing each entity pair for each entity pair, and perform a final relationship prediction based on the weighted instance representation.

[0234] Specifically, in order to deal with the label noise problem caused by remote supervision, the graph-enhanced relationship extraction module of this embodiment adopts a multi-instance learning method based on the attention mechanism:

[0235] a) Instance-level attention: For each entity pair, the attention weights of all sentences containing the entity pair are calculated.

[0236] b) Bag-level prediction: The final relation prediction is made based on the weighted instance representation.

[0237] As a specific example, the denoising process of the multi-instance learning method can be expressed as:

[0238]

[0239] in, It is The expression of a sentence, is the relation prediction for the sentence, is the attention weight.

[0240] The multi-task classifier of this embodiment can perform relationship classification and entity type prediction simultaneously, which greatly improves the generalization ability of the model in the graph enhanced relationship extraction module.

[0241] In this embodiment, the training and optimization process of the graph enhanced relation extraction module (GARE) is trained using a medical text dataset with relation annotations.

[0242] Specifically, the graph-enhanced relation extraction module (GARE) adopts a multi-task learning framework, and the loss function includes relation classification loss, entity type refinement loss and graph structure prediction loss:

[0243]

[0244] in, is the cross entropy loss for relation classification, is the cross entropy loss for entity type refinement, is the loss of graph structure prediction (such as edge existence prediction), and is the weight coefficient for balancing different tasks.

[0245] The output of the Graph-Enhanced Relation Extraction module (GARE) will be formatted into a new data structure called "Relation-Enhanced Text Representation" (RETR):

[0246] As a specific example, RETR = {

[0247] 'original_text': str,

[0248] 'cleaned_text': str,

[0249] 'sentences': List[Dict{

[0250] 'text': str,

[0251] 'tokens': List[str],

[0252] 'entities': List[Dict{

[0253] 'text': str,

[0254] 'start': int,

[0255] 'end': int,

[0256] 'type': str,

[0257] 'refined_type': str

[0258] }],

[0259] 'relations': List[Dict{

[0260] 'head': int,

[0261] 'tail': int,

[0262] 'type': str,

[0263] 'confidence': float

[0264] }],

[0265] 'graph': Dict{

[0266] 'nodes': List[int],

[0267] 'edges': List[Tuple[int, int, str]]

[0268] }

[0269] }]

[0270] }

[0271] Based on EETR, this RETR structure adds a 'relations' field to each sentence, which contains the extracted entity relationship information, and a 'graph' field, which records the constructed graph structure information. At the same time, a 'refined_type' field is added to the entity information to indicate the refined entity type.

[0272] The Graph Enhanced Relation Extraction Module (GARE) of this embodiment adopts the innovative "Hierarchical Graph Attention Network" (HGAT) architecture, which combines the powerful feature extraction capability of the pre-trained language model, the structural perception capability of the graph neural network, and the generalization capability of multi-task learning.

[0273] In this embodiment, the large model-assisted verification and optimization module, namely Large Language Model-based Verification and Optimization, is referred to as LLMVO. As an innovative component of the medical intelligent knowledge graph automatic construction system described in the embodiment of the present invention, it aims to verify and optimize the relational enhanced text representation (RETR) by utilizing the powerful semantic understanding and reasoning capabilities of large-scale language models. The LLMVO mainly focuses on entity consistency checking, relationship rationality verification, and knowledge graph enhancement.

[0274] In some specific embodiments, the large model-assisted verification and optimization module (LLMVO) is specifically configured to perform the following:

[0275] 1. Input processing

[0276] Define a function f: RETR → P to convert RETR into a prompt suitable for large model processing :

[0277]

[0278] in, is context information, yes The textual representation of Is a specific query or instruction.

[0279] At the same time, a dynamic prompt generation mechanism is introduced:

[0280] (a) Context-aware: Dynamically adjust context information based on the content of RETR .

[0281]

[0282] (b) Adaptive query generation: Automatically generate the most suitable query based on the verification task .

[0283]

[0284] (c) Diversified prompt templates: Use multiple predefined prompt templates and select the most appropriate one according to the task.

[0285]

[0286] In summary, the input processing function can be expressed as:

[0287]

[0288] 2. Verification process

[0289] make The output of the large model for the prompt. Define three post-processing functions and :

[0290] a) Entity consistency check: in is a collection of entities, Is Entity The consistency score.

[0291] b) Verification of relationship rationality: in is a set of relations, It's a relationship rationality score.

[0292] c) Knowledge completion and reasoning: in is the new knowledge set obtained by reasoning, It is knowledge Confidence score of .

[0293] The verification process is as follows:

[0294] (1) Multi-round verification: For entities and relations with lower confidence, multiple rounds of verification are performed, using different prompts in each round.

[0295]

[0296] (2) Cross-validation: Use entity information to verify relationships, and use relationship information to verify entities to improve overall consistency.

[0297]

[0298] (3) Knowledge graph reasoning: Based on existing entities and relationships, infer possible new relationships or entities.

[0299]

[0300] 3. Output Generation

[0301] Defining the optimization function :

[0302]

[0303] The formal representation is:

[0304] in:

[0305]

[0306] is a collection of entity types, is a set of new knowledge types, and are the optimized entity and relationship sets, respectively. It is a newly added collection of knowledge. and Is a binary verification flag that indicates whether the entity or relationship has passed verification.

[0307] Introducing decision-making mechanisms and interpretability analysis:

[0308] Threshold optimization: Automatically adjust the acceptance / rejection threshold using machine learning methods.

[0309] Conflict Resolution: When entity or relationship conflicts arise, they are resolved using a rule-based system.

[0310] Explainability Reports: Generate concise explanations for each validation and inference decision.

[0311] Final output:

[0312]

[0313] In this embodiment, the end-to-end differential learning framework, namely the End-to-End Differentiable Learning Framework, referred to as E2DL, is designed to seamlessly integrate the hierarchical multi-task entity recognition module (HMT-NER), the graph enhanced relationship extraction module (GARE) and the large model assisted verification and optimization module (LLMVO) to achieve differentiable optimization of the entire process.

[0314] The core idea of ​​the E2DL framework is to achieve joint optimization of the entire knowledge graph construction process through end-to-end back-propagation. The framework converts the discrete outputs of each module into continuous soft decisions, so that the gradient can be smoothly propagated throughout the system.

[0315] In some specific embodiments, the end-to-end differential learning framework (E2DL) is specifically configured to achieve the following:

[0316] 1. Differentiable connections between modules

[0317] a) Differentiable connection between the Hierarchical Multi-Task Entity Recognition module (HMT-NER) and the Graph Enhanced Relation Extraction module (GARE):

[0318] Leveraging BioBERT output from the Hierarchical Multi-Task NER module (HMT-NER) , define the soft entity representation function:

[0319]

[0320] in, and is a learnable parameter, is the sigmoid function.

[0321] b) Differentiable connection from the Graph Enhanced Relation Extraction module (GARE) to the Large Model Assisted Validation and Optimization module (LLMVO):

[0322] Leveraging the graph attention network output of the Graph Enhanced Relation Extraction module (GARE)

[0323]

[0324] Define the soft relation representation function:

[0325]

[0326] in, and are learnable parameters.

[0327] c) Differentiable knowledge graph construction: Defining a differentiable knowledge graph for:

[0328]

[0329] Among them, the confidence score The calculation is:

[0330]

[0331] d) Differentiable integration of Large Model Assisted Verification and Optimization (LLMVO) module:

[0332] Convert the output of the LLMVO module into a differentiable validation score:

[0333]

[0334] in, It is a large language model. and are prompts for entities and relationships, respectively. is a learnable validation weight.

[0335] 2. Unified end-to-end loss function

[0336] Configure a unified end-to-end loss function:

[0337]

[0338] in,

[0339]

[0340] 3. Optimization strategy

[0341] a) SAEA: To solve the problem of inconsistent entity boundaries between HMT-NER and GARE, the SAEA mechanism is introduced:

[0342]

[0343] in, , is the alignment scoring function.

[0344] Relation-Aware Entity Optimization (RAER): Reverse optimization of entity representation using relation information:

[0345]

[0346] in, Is with entity All related relationships, is a learnable fine-tuning matrix, is the fine-tuning strength coefficient.

[0347] b) Large Model Guided Noise Contrastive Learning (LGCL): LGCL is introduced to improve the discriminability of representations:

[0348] in, is the current representation, is a representation verified by LLMVO to be correct, is a negative sample, is the similarity function, is the temperature parameter.

[0349] c) Adopting an alternating optimization strategy, in each iteration:

[0350] (1) Fix LLMVO, optimize HMT-NER, GARE and connection modules:

[0351] ;

[0352] (2) Fix other modules and optimize LLMVO:

[0353]

[0354] in, is the learning rate, which can be adjusted using the adaptive optimizer Adam.

[0355] The end-to-end differential learning framework (E2DL) of this embodiment realizes the full differentiability of the entire knowledge graph construction process by converting discrete decisions into continuous soft decisions. It not only retains the advantages of each module, but also enhances the synergy between modules through a unified loss function and innovative optimization strategies. Therefore, the full differentiable integration based on the end-to-end differential learning framework (E2DL) can significantly improve the overall performance of the system and provide a more unified, efficient and robust solution for the automatic construction of medical text knowledge graphs.

[0356] An embodiment of the present invention provides an automatic construction system of a medical intelligent knowledge graph based on end-to-end differential learning. Through an innovative end-to-end differential learning framework (E2DL), a hierarchical multi-task entity recognition module (HMT-NER), a graph enhanced relationship extraction module (GARE) and a large model assisted verification and optimization module (LLMVO) are integrated therein in a differentiable manner, thereby forming a closed-loop knowledge extraction-verification-optimization process, and realizing a full-process differentiable, automated and precise conversion from unstructured medical texts to high-quality knowledge graphs.

[0357] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0358] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0359] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0360] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0361] It should be noted that in the claims, any reference signs placed between brackets shall not be construed as limiting the claims. The word "comprising" does not exclude the presence of components or steps not listed in the claim. The word "a" or "an" preceding a component does not exclude the presence of a plurality of such components. The invention may be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In a unit claim enumerating several means, several of these means may be embodied by the same item of hardware. The use of the words first, second, and third etc. does not indicate any order. These words may be interpreted as names.

[0362] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0363] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

[0364] In the description of the present invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0365] In the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", "fixed" and the like should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0366] In the present invention, unless otherwise clearly specified and limited, a first feature being "above" or "below" a second feature may mean that the first and second features are in direct contact, or the first and second features are in indirect contact through an intermediate medium. Moreover, a first feature being "above", "above" or "above" a second feature may mean that the first feature is directly above or obliquely above the second feature, or simply means that the first feature is higher in level than the second feature. A first feature being "below", "below" or "below" a second feature may mean that the first feature is directly below or obliquely below the second feature, or simply means that the first feature is lower in level than the second feature.

[0367] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms should not be understood as necessarily being directed to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, unless they are contradictory.

[0368] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. A medical intelligent knowledge graph automatic construction system based on end-to-end differential learning, characterized by: It includes an end-to-end differential learning framework and a hierarchical multi-task entity recognition module, a graph-enhanced relationship extraction module, and a large model-assisted verification and optimization module integrated in the end-to-end differential learning framework; the hierarchical multi-task entity recognition module, the graph-enhanced relationship extraction module, and the large model-assisted verification and optimization module are sequentially differentiably connected; The hierarchical multi-task entity recognition module includes a BioBERT encoder, a bidirectional GRU layer, an attention mechanism, a hierarchical classifier, and a conditional random field layer connected in sequence, and is used to capture deep semantic features, capture long-term dependency relationships, dynamically assign different weights to different parts of the input sequence according to the context, perform hierarchical classification, and calculate the optimal label combination of the sequence according to the input pre-processed medical text sequence, so as to identify and extract various medical entities; The graph-enhanced relationship extraction module includes a context encoder, a graph structure, a graph attention network, and a multi-task classifier connected in sequence, which are used to capture contextual semantic information processing, construct a sentence-level graph structure processing, capture semantic relationship processing at different levels through a multi-layer graph attention mechanism, and perform relationship classification and entity type prediction processing according to the input medical entity, so as to identify and extract the semantic relationship between medical entities and output relationship-enhanced text; The large model-assisted verification and optimization module is used to verify and optimize the relationship-enhanced text representation using a large-scale language model.

2. The medical intelligent knowledge graph automatic construction system based on end-to-end differential learning according to claim 1, characterized in that: The BioBERT encoder includes: The first multi-layer Transformer encoder, each layer of which has a self-attention mechanism and a feed-forward neural network; and The context-sensitive representation unit is used to capture long-distance dependencies in the sequence through the self-attention mechanism and generate context-sensitive word representations.

3. The medical intelligent knowledge graph automatic construction system based on end-to-end differential learning according to claim 1, characterized in that: The bidirectional GRU layer includes: A bidirectional processing unit for processing input sequences bidirectionally and in parallel, capturing long-range dependencies in the sequences and generating context-sensitive word representations; and The gate control mechanism unit includes an update gate and a reset gate, and is used to dynamically control the flow of information during the processing of the bidirectional processing unit.

4. The medical intelligent knowledge graph automatic construction system based on end-to-end differential learning according to claim 1, characterized in that: The attention mechanism includes: A multi-angle feature extraction unit, which uses multiple attention heads to extract different features of the input sequence at the same time; The attention weight calculation unit is used to dynamically calculate the weights of different features according to the context through multi-head splicing and linear transformation.

5. The medical intelligent knowledge graph automatic construction system based on end-to-end differential learning according to claim 1, characterized in that: The conditional random field layer includes: A transition probability matrix unit, used to learn the transition probability between labels and obtain a transition matrix; and The emission probability unit is used to combine the output of the hierarchical classifier and the transfer matrix to calculate the optimal label sequence.

6. The medical intelligent knowledge graph automatic construction system based on end-to-end differential learning according to claim 1, characterized in that: The context encoder comprises The second multi-layer Transformer encoder is used to generate context-sensitive word representations; An attention pooling unit is used to pool the output of the second multi-layer Transformer encoder using an attention mechanism to obtain a sentence-level representation.

7. The medical intelligent knowledge graph automatic construction system based on end-to-end differential learning according to claim 1, characterized in that: The graph attention network includes multiple graph attention layers; Each of the multiple graph attention layers adopts residual connection, and each graph attention layer adopts a multi-head attention mechanism; wherein the attention calculation process of the multi-head attention mechanism will consider the type of edge.

8. The medical intelligent knowledge graph automatic construction system based on end-to-end differential learning according to claim 1, characterized in that: The multi-task classifier adopts a multi-instance learning method based on an attention mechanism to calculate the attention weights of all sentences containing each entity pair for each entity pair, and perform final relationship prediction based on the weighted instance representation.

9. The medical intelligent knowledge graph automatic construction system based on end-to-end differential learning according to claim 1, characterized in that: It also includes a preprocessing module; the preprocessing module is integrated into the end-to-end differential learning framework and connected to the input end of the hierarchical multi-task entity recognition module; The preprocessing module is used to convert the original unstructured medical text into a standardized, structured preprocessed medical text sequence.

10. The medical intelligent knowledge graph automatic construction system based on end-to-end differential learning according to claim 9, characterized in that: The preprocessing module includes a text cleaning unit, a sentence segmentation unit, a word standardization unit, a part-of-speech tagging unit and a dependency syntax analysis unit.

Images