Medical word sense disambiguation method based on kernel homomorphism decoupling and riemann funnel

Abstract

This invention relates to a medical semantic disambiguation method based on kernel homomorphic decoupling and Riemannian funnel. The invention first processes the biomedical corpus by segmenting the text containing ambiguous words, tagging parts of speech, semantic classes, and inserting target markers, mapping the processed training and testing data. Then, the self-catalytic cognitive disambiguation network is trained using the training data to obtain an optimized model. Based on the optimized model, inference is performed on the testing data, calculating the confidence distribution and nonlinear cognitive state uncertainty of the ambiguous words under each semantic category. Rejection is triggered by an uncertainty threshold; for safe samples that are not rejected, the semantic category with the highest confidence is the final semantic category of the ambiguous word. This invention effectively removes background noise from long texts and prevents overfitting, while constructing a highly reliable state collapse rejection defense, enabling safer and more accurate determination of the true meaning of ambiguous biomedical words.

Description

Technical Field

[0001] This invention relates to a medical semantic disambiguation method based on nuclear homomorphic decoupling and Riemann funnel, which has good applications in the fields of biomedical natural language processing, electronic medical record structuring, and intelligent assisted diagnosis. Background Technology

[0002] In the field of biomedical natural language processing, medical vocabulary exhibits strong polysemy. Accurately determining its semantics in a specific medical context is the core foundation for medical information retrieval, intelligent consultation, and clinical decision support. Since the accuracy of word meaning disambiguation is directly related to the safety and reliability of subsequent clinical diagnosis, the academic community has widely introduced deep learning algorithms such as convolutional neural networks (CNN) and pre-trained language models (such as BERT) in recent years for semantic feature extraction and classification. These methods utilize massive corpora to learn the contextual representation of words, significantly improving disambiguation performance in general scenarios, and have become the mainstream technical path in the field of biomedical natural language processing.

[0003] However, existing deep learning algorithms still face severe challenges in practical medical applications: First, long medical records contain a large amount of redundant background noise, which leads to severe coupling of the extracted disambiguation features and weak anti-interference ability. Second, for small samples of rare diseases under the long-tail distribution of medical data, the "static point features" generated by traditional models lack dynamic representation and are prone to overfitting, which limits the generalization ability of the model. Most importantly, existing algorithms generally have the defect of "blind confidence". Even when faced with extremely difficult samples with missing key contexts, they will still force the output of prediction results. They lack the rejection boundary and cognitive uncertainty monitoring mechanism required to ensure medical safety and are unable to meet the reliability requirements of high-risk clinical scenarios. Summary of the Invention

[0004] To address the issues of severe noise in long biomedical texts, overfitting in small samples of rare diseases, and the lack of medical safety boundaries due to the overconfidence of traditional classification algorithms, this invention discloses a medical semantic disambiguation method based on kernel homomorphic decoupling and Riemann funnel.

[0005] Therefore, the present invention provides the following technical solution:

[0006] 1. A medical semantic disambiguation method based on nuclear homomorphic decoupling and Riemannian funnel, for ambiguous words. have semantic categories The method is characterized by comprising the following steps:

[0007] Step 1: Perform lexical segmentation, part-of-speech tagging, and semantic category tagging on the entire biomedical WSD corpus, obtain the part-of-speech tag sequence and semantic category sequence corresponding to each word, and locate ambiguous words. Target markers are inserted before and after the text to form a sequence of marked texts with an attention-guided mechanism. The processed corpus is then divided into training corpus and test corpus.

[0008] Step 2: For the labeled text sequences, part-of-speech tag sequences, and semantic category sequences extracted from the training corpus of the biomedical WSD corpus, use the word segmenter of the pre-trained language model BERT to perform dictionary index mapping and convert them into input tensor sequences as training data. For the labeled text sequences, part-of-speech tag sequences, and semantic category sequences extracted from the test corpus of the biomedical WSD corpus, use the word segmenter of the pre-trained language model BERT to perform dictionary index mapping and convert them into input tensor sequences as test data.

[0009] Step 3: Optimize the autocatalytic cognitive disambiguation network using training data to obtain the optimized autocatalytic cognitive disambiguation network. The autocatalytic cognitive disambiguation network is a model based on fractional-order kernel homomorphic decoupling and anisotropic Riemannian information funnel, which includes a BERT feature extractor, a feature manifold splitting network, a fractional-order regenerating kernel homomorphic decoupling module, an anisotropic Riemannian information funnel module, and a Laplace evidence reasoning layer.

[0010] Step 4: The testing process, namely semantic classification and state collapse determination, involves inputting test data into the optimized autocatalytic cognitive disambiguation network and calculating ambiguous words. The confidence distribution and nonlinear cognitive state uncertainty under each semantic category are used to trigger rejection decisions through an uncertainty threshold. Among them, for safe samples that are not rejected, the semantic category with the highest confidence is the ambiguous word. The final semantic category.

[0011] 2. The medical semantic disambiguation method based on nuclear homomorphic decoupling and Riemann funnel according to claim 1, characterized in that, in step 1, the specific steps for preprocessing and segmenting the entire biomedical WSD corpus are as follows:

[0012] Step 1-1: Use the text processing module to read Chinese sentences and corresponding ambiguous words from the biomedical corpus. The medical professional lexicon and natural language processing toolkit were used to perform word segmentation, part-of-speech tagging and semantic category tagging on the Chinese sentences to obtain a part-of-speech tag sequence and semantic category sequence corresponding to each word.

[0013] Steps 1-2: Locating ambiguous words in Chinese sentences And insert special markers [TGT] and [ / TGT] before and after it respectively to form a marked text sequence with an attention guidance mechanism;

[0014] Steps 1-3 divide the entire corpus, which includes the above-mentioned marked text sequences, part-of-speech tag sequences, and semantic category sequences, into training corpus and test corpus.

[0015] 3. The medical semantic disambiguation method based on nuclear homomorphic decoupling and Riemann funnel according to claim 1, characterized in that, in step 2, the construction of training data and test data specifically includes the following steps:

[0016] Step 2-1: For the divided training and test corpora, use the Tokenizer of the pre-trained language model BERT to perform word segmentation and dictionary index mapping on the marked text sequences to obtain discrete Token ID sequences.

[0017] Step 2-2 involves uniformly truncating and padding the Token ID sequence to generate the input sequence tensor input_ids and the corresponding attention mask tensor attention_mask;

[0018] Steps 2-3 extract the input_ids and attention_mask tensors corresponding to the training corpus, align them with their true semantic category labels, and convert them into PyTorch Dataset format;

[0019] Steps 2-4 use DataLoader to encapsulate the training Dataset into iterable training data;

[0020] Steps 2-5 extract the input_ids and attention_mask tensors corresponding to the test corpus and convert them into PyTorch Dataset format;

[0021] Steps 2-6 use DataLoader to encapsulate the test Dataset into iterable test data.

[0022] 4. The medical semantic disambiguation method based on nuclear homomorphic decoupling and Riemann funnel according to claim 1, characterized in that, in step 3, training and optimizing the autocatalytic cognitive disambiguation network specifically involves the following steps:

[0023] Step 3-1: Input the training data into the initialized autocatalytic cognitive disambiguation network and use the feature extractor to extract the pooling layer output features of the text.

[0024] Step 3-2 inputs the corresponding part-of-speech tag sequence and semantic category sequence into independent embedding layers to map and obtain part-of-speech feature vectors and semantic category feature vectors. Then, the pooling layer output features, part-of-speech feature vectors, and semantic category feature vectors are multi-source heterogeneous concatenation and fusion, and the fused features are used as the initial homeomorphic manifold features. ;

[0025] Step 3-3 will define the initial homeomorphic manifold features. The input is fed into a feature manifold splitting network, where parallel fully connected layers combined with the GELU activation function are used to map the features into semantic backbone manifold features. and background fluctuation manifold characteristics The mapping process is as follows:

[0026]

[0027]

[0028] in, , This is the weight matrix. , It is the bias vector;

[0029] Steps 3-4 will and The input is fed into the fractional-order regeneration kernel homomorphic decoupling module, which constructs feature correlation matrices based on the exponential distance between feature vectors. and The calculation process for the matrix elements is as follows:

[0030]

[0031] in, For bandwidth parameters, and The first one extracted from the current training data is the second one. The and the first 1 feature vector sample;

[0032] Steps 3-5 calculate the fractional-order decoupling penalty error based on the feature correlation matrix. The calculation process is as follows:

[0033]

[0034] in The total number of samples in the training data. Represents the trace operation of a matrix. The exponent of the fractional homomorphism and It is used to perform fractional-order damping attenuation of the coupling degree of nonlinear features in high-dimensional space. The centering matrix is ​​used to map features to a high-dimensional kernel projection space with zero mean. The calculation process of this centering matrix is ​​as follows:

[0035]

[0036] in, for The identity matrix, For length is A column vector of all 1s;

[0037] Steps 3-6 will Input is fed into the anisotropic information funnel module to predict the high-dimensional feature sphere. The direction mean vector on Concentration parameters and covariance deformation tensor The posterior probability density function of the feature is calculated using the prediction parameters. Its functional expression is as follows:

[0038]

[0039] in, For feature dimension, This is a thermodynamic distribution temperature scaling factor used to smooth the probability density surface of the manifold space in the early stages of training to prevent local gradient collapse. It is a high-dimensional volume constraint operator used to analyze the nonlinear mapping relationship between parameters and high-dimensional spatial volume, ensuring that the constructed posterior distribution satisfies the geometric constraint that the probability integral is always 1;

[0040] Step 3-7 Calculate the probability density function With the target prior uniform distribution Feature compression regularization loss The loss is the distribution divergence between the two, and the calculation process is as follows:

[0041]

[0042] in, For high-dimensional characteristic spheres The prior uniform probability density on, The dynamic information capacity control threshold is used to preserve the basic semantic representation capacity during training and prevent semantic collapse caused by excessive compression of the feature manifold.

[0043] Steps 3-8 combined , and Generalization features are obtained by performing spherical constrained noisy sampling. The sampling process is as follows:

[0044]

[0045] in It is a random noise matrix. This is a manifold curvature adaptive perturbation damping system used to dynamically adjust the injection intensity of noise based on the local curvature of a high-dimensional characteristic sphere. This is a spatial distance normalization operation. It uses Euclidean distance to force a strict projection and constraint of the features after adding local noise perturbation onto a given surface. On the manifold surface of the 3D feature sphere, thereby ensuring that the generalization feature satisfies the geometric prerequisites of the probability density function;

[0046] Steps 3-9 will generalize features Input evidence reasoning layer, output ambiguous words In each semantic category The amount of non-negative evidence below The calculation process is as follows:

[0047]

[0048] in, and These are the Laplace evidence reasoning layers corresponding to the first... Weight vectors and biases for each semantic category;

[0049] Steps 3-10: Based on the amount of non-negative evidence Construct evidence distribution parameters and total amount of evidence And calculate the expected classification error. The calculation process is as follows:

[0050]

[0051]

[0052]

[0053] in, This is a long-tail category sensitivity compensation coefficient, used to adaptively amplify the weight of the medical rare disease category in gradient backpropagation. For topological focusing index, This is a logarithmic derivative operator used to calculate the logarithmic expectation of the polynomial prediction probability under the current evidence parameters. It smoothly maps the discrete classification error to a continuous distribution space, measuring the expected error of the model in the cognitive blind spot. One-hot encoding for the real label;

[0054] Step 3-11 Based on distribution parameters Calculate the logarithmic divergence penalty used to punish overconfidence. The calculation process is as follows:

[0055]

[0056] in, This is a continuous factorial extension operator used to analytically compute the multivariate integral normalization constant of the evidence distribution. It measures the global probability difference between the model's predicted distribution and the prior uniform distribution in the divergence integral term. It is a nonlinear truncation operator. To provide cognitive safety divergence margin, a reasonable feature ambiguity is preserved in the penalty process to prevent the model from losing its basic diagnostic intuition for similar clinical symptoms due to over-regularization. The formula above... The calculation process is as follows:

[0057]

[0058] in, Correct parameters for the mask;

[0059] Step 3-12 will , , and The total functional error is obtained by weighted summation. The calculation process is as follows:

[0060]

[0061] in, , and The hyperparameter weights are based on the total functional error. Backpropagation is performed to update the network parameters layer by layer until the optimized autocatalytic cognitive disambiguation network is obtained.

[0062] 5. The medical semantic disambiguation method based on nuclear homomorphic decoupling and Riemann funnel according to claim 1, characterized in that, in step 4, the test data is subjected to inference and state collapse determination, the specific process of which is as follows:

[0063] Step 4-1: Input the test data into the optimized autocatalytic cognitive disambiguation network, which then passes through the BERT feature extractor and the feature manifold splitting network to extract the semantic backbone manifold features. ;

[0064] Step 4-2 will Input anisotropic Riemannian information funnel module to predict the direction mean vector During the testing phase, reparameterization sampling is stopped, and the direction mean vector is directly applied. As a generalization feature, it is input into the Laplace evidence reasoning layer;

[0065] Step 4-3 Extract the amount of non-negative evidence output from the Laplace evidence reasoning layer. Calculate the evidence distribution parameters and total amount of evidence Then calculate ambiguous words. In semantic categories Confidence level The calculation process is as follows:

[0066]

[0067]

[0068]

[0069] in The total number of semantic categories;

[0070] Step 4-4 Based on parameters and Calculate the uncertainty of nonlinear cognitive state The calculation process is as follows:

[0071]

[0072] in, The state collapse sensitivity index and It is used to steeply amplify the cognitive uncertainty in the high information entropy state in the feature space of extremely difficult medical records, thereby accelerating the triggering of the state collapse rejection mechanism in high-risk clinical scenarios.

[0073] Steps 4-5: Construct the state determination condition function To reduce the uncertainty of nonlinear cognitive states With the preset safety barrier threshold The comparison and segmented output of the judgment results are performed. The calculation process of the judgment condition function is as follows:

[0074]

[0075] Among them, when At that time, the current test data is judged to be an extremely difficult sample and is directly pushed into the manual review queue. When the current test data is determined to be a safe sample, the output with the highest confidence level is given. semantic categories As an ambiguous word The final predicted category.

[0076] Beneficial effects:

[0077] 1. This invention effectively overcomes the limitation of insufficient representation of single features in complex contexts by multi-source heterogeneous fusion of deep context extracted by pre-trained language models with part-of-speech tags and semantic class rules containing medical prior knowledge, thereby extracting initial homeomorphic manifold features with extremely high representational ability.

[0078] 2. By utilizing the feature manifold splitting network and the fractional-order regenerating kernel homomorphic decoupling module, redundant background noise in long medical records can be forcibly removed. By introducing a fractional-order homomorphic mapping exponent to dampen the coupling of nonlinear features in high-dimensional space, the purity of the extracted semantic backbone manifold features is ensured, which greatly improves the model's anti-interference ability when facing long medical texts.

[0079] 3. This invention innovatively introduces an anisotropic Riemannian information funnel module, which, while smoothing the probability density surface of the manifold space, implicitly completes dynamic data enhancement at the feature level through an adaptive noise-adding sampling mechanism on a high-dimensional feature sphere. This fundamentally solves the problem of model overfitting caused by small sample data of rare diseases under long-tailed distribution in the medical field.

[0080] 4. In the test and reasoning phase, this invention completely solves the problem of blind confidence in traditional classification algorithms by constructing a safety rejection mechanism based on nonlinear cognitive state uncertainty. When encountering extremely difficult samples with extremely lacking contextual information, the system uncertainty will show a steep amplification trend and accurately trigger the interception and verification logic, which significantly improves the safety bottom line control capability of this method in clinical practice. Attached Figure Description

[0081] Figure 1 An overall flowchart of a medical semantic disambiguation method provided in this embodiment of the invention.

[0082] Figure 2 The flowchart of training and error optimization of the self-catalytic cognitive disambiguation network provided in this embodiment of the invention.

[0083] Figure 3 The flowchart for state collapse determination and rejection during the testing phase provided in this embodiment of the invention Detailed Implementation

[0084] To ensure that the technical solutions in the embodiments of the present invention can be clearly and completely described, further detailed descriptions are provided below in conjunction with the embodiments.

[0085] Taking the disambiguation of the ambiguous word "leukoplakia" in the real biomedical medical record text "The patient's oral mucosa shows leukoplakia, accompanied by mild pain and ulceration, and the possibility of cancer is ruled out" as an example, in this embodiment, the semantic categories of the ambiguous word "leukoplakia" in the corpus are three (i.e. ), respectively (Oral leukoplakia) (Vulvar leukoplakia) (Vitiligo).

[0086] The training and testing process of the self-catalytic cognitive disambiguation network in this embodiment of the invention includes the following steps:

[0087] Step 1: Perform lexical segmentation, part-of-speech tagging, and semantic category tagging on the entire biomedical WSD corpus, obtain the part-of-speech tag sequence and semantic category sequence corresponding to each word, and locate ambiguous words. Target markers are inserted before and after to form a marked text sequence with an attention-guided mechanism. The processed corpus is then divided into training and testing corpora. The specific steps are as follows:

[0088] Step 1-1: Use the text processing module to read Chinese sentences and corresponding ambiguous words from the biomedical corpus. The system utilizes a medical lexicon and a natural language processing toolkit to perform lexical segmentation, part-of-speech tagging, and semantic category tagging on the Chinese sentence, obtaining a part-of-speech tag sequence and semantic category sequence corresponding to each word. In this embodiment, the read Chinese sentence is "The patient's oral mucosa shows leukoplakia, accompanied by mild pain and ulceration, and the possibility of cancer is ruled out." The corresponding ambiguous words... For the term "leukoplakia", after word segmentation and part-of-speech and semantic tagging, the obtained sequence portion is truncated as: "Patient / Person Oral cavity / BodyPart Mucosa / Tissue Appears / Action Leukoplakia / Disease...";

[0089] Steps 1-2: Locating ambiguous words in Chinese sentences The special markers [TGT] and [ / TGT] are inserted before and after the text to form a marked text sequence with an attention guidance mechanism. In this embodiment, the marker module is used to locate the ambiguous word "leukoderma" and special markers are inserted before and after it. The marked text sequence formed after conversion is partially extracted as: "Patient / Person Oral / BodyPart Mucosa / Tissue Appears / Action [TGT]Leukoderma / Disease[ / TGT]...";

[0090] Steps 1-3 divide the entire corpus containing the above-mentioned marked text sequence, part-of-speech tag sequence, and semantic category sequence into training corpus and test corpus. In this embodiment, the entire biomedical WSD corpus is randomly divided according to a preset ratio. The samples containing the above-processed "white spot" sentences are divided into the training corpus as part of the training data, and the remaining corpus is divided into the training corpus and the test corpus respectively.

[0091] Step 2: For the labeled text sequences, part-of-speech tag sequences, and semantic category sequences extracted from the training corpus of the Biomedical WSD corpus, the dictionary index mapping is performed using the tokenizer of the pre-trained language model BERT, and the sequences are converted into input tensor sequences as training data. For the labeled text sequences, part-of-speech tag sequences, and semantic category sequences extracted from the test corpus of the Biomedical WSD corpus, the dictionary index mapping is performed using the tokenizer of the pre-trained language model BERT, and the sequences are converted into input tensor sequences as test data. The specific process is as follows:

[0092] Step 2-1: For the divided training and test corpora, the tokenizer of the pre-trained language model BERT is used to perform word segmentation and dictionary index mapping on the marked text sequences to obtain discrete token ID sequences. In this embodiment, the marked text sequences of the training and test corpora obtained in Step 1-2 are input, and the tokenizer of the pre-trained language model BERT is used to perform word segmentation and dictionary entry index mapping on them, thereby extracting the original numerical sequences of discrete token IDs that have not been padded or truncated.

[0093] Step 2-2 performs uniform length truncation and padding on the Token ID sequence to generate the input sequence tensor input_ids and the corresponding attention mask tensor attention_mask. In this embodiment, the discrete Token ID sequence is normalized by combining the uniform length truncation and padding rule of 128, and the insufficient part at the end is padded with 0, generating a dimension of The input sequence tensor input_ids is used to generate a corresponding binary attention mask tensor attention_mask. The truncated content is displayed as follows: input_ids = [101,2642,5442,1366,5579,7945,5606,1139,4385,138,…,0,0], attention_mask = [1,1,1,1,1,…,1,1,1,1,…,0,0];

[0094] Steps 2-3 extract the input_ids and attention_mask tensors corresponding to the training corpus, align them with their true semantic category labels, and convert them to PyTorch Dataset format. In this embodiment, all input_ids and attention_mask tensors generated from the training corpus samples are extracted and aligned with the true one-hot encoded semantic category labels (e.g., for the ambiguous word "white spot," its true semantic category is...). When, label alignment is represented as The data is then precisely aligned and correlated along the sample dimension, and then uniformly packaged and converted into the standard Dataset data format of the PyTorch framework.

[0095] Steps 2-4 use DataLoader to encapsulate the training dataset into iterable training data. In this embodiment, the built-in DataLoader component of PyTorch is called to encapsulate the training dataset constructed above, and the training batch size is set to 16. The parameters for shuffling the dataset are configured to transform it into batch training data that can provide iterable tensors in batches.

[0096] Steps 2-5 extract the input_ids and attention_mask tensors corresponding to the test corpus and convert them into PyTorch's Dataset format. In this embodiment, all input_ids tensors and attention_mask tensors corresponding to the test corpus samples are extracted. Without including the real semantic category labels, they are directly encapsulated and converted into the standard test Dataset data format of the PyTorch framework.

[0097] Steps 2-6 utilize DataLoader to encapsulate the test Dataset into iterable test data. In this embodiment, the DataLoader component is also used to encapsulate the test Dataset, set the corresponding reading parameters, and maintain the original order of the samples during the testing phase. Finally, it is transformed into iterable batch test data that the network can call in batches during the test inference phase.

[0098] Step 3: Optimize the autocatalytic cognitive disambiguation network using training data to obtain the optimized autocatalytic cognitive disambiguation network. The autocatalytic cognitive disambiguation network is a model based on fractional-order kernel homomorphic decoupling and anisotropic Riemannian information funnel, comprising a BERT feature extractor, a feature manifold splitting network, a fractional-order regenerating kernel homomorphic decoupling module, an anisotropic Riemannian information funnel module, and a Laplace evidence reasoning layer. In this embodiment, the forward propagation and feature manifold extraction processes are performed before the training phase. The specific steps are as follows:

[0099] Step 3-1: Input the training data into the initialized autocatalytic cognitive disambiguation network and use the feature extractor to extract the pooling layer output features of the text. In this embodiment, the batch training data containing input_ids and attention_mask is input into the initialized autocatalytic cognitive disambiguation network. After forward propagation calculation by the pre-trained language model BERT feature extractor, the output features of the batch of medical record text in the pooling layer (Pooled Output) are extracted.

[0100] Step 3-2 inputs the corresponding part-of-speech tag sequence and semantic category sequence into independent embedding layers to map and obtain part-of-speech feature vectors and semantic category feature vectors. Then, the pooling layer output features, part-of-speech feature vectors, and semantic category feature vectors are multi-source heterogeneous concatenation and fusion, and the fused features are used as the initial homeomorphic manifold features. In this embodiment, the corresponding part-of-speech tag sequences such as " / Disease" and semantic category sequences are input into independent embedding layers of the network to map and obtain corresponding part-of-speech and semantic category feature vectors. Then, the features output from the pooling layer are multi-source heterogeneous concatenation and fusion with the above feature vectors to obtain an 896-dimensional initial homeomorphic manifold feature that condenses the global context and medical prior knowledge. ;

[0101] Step 3-3 will define the initial homeomorphic manifold features. The input is fed into a feature manifold splitting network, where parallel fully connected layers combined with the GELU activation function are used to map the features into semantic backbone manifold features. and background fluctuation manifold characteristics In this embodiment, the aforementioned 896 dimensions are... The input is fed into a feature manifold splitting network, which uses parallel fully connected layers to reduce dimensionality and remove parts of the manifold. The mapping process is as follows:

[0102]

[0103]

[0104] in, , The dimensions set are respectively Independent weight matrix, , The bias vector, after extraction, outputs the dimension. Core semantic backbone manifold features (For example, after extraction, the first three elements of its tensor are) This is accompanied by the generation of 384-dimensional background fluctuation manifold features. (For example, the first three elements of its tensor are) );

[0105] Steps 3-4 will and The input is fed into the fractional-order regeneration kernel homomorphic decoupling module, which constructs feature correlation matrices based on the exponential distance between feature vectors. and In this embodiment, the bandwidth parameter is set. The feature correlation matrices mentioned above are constructed based on the exponential distance between feature vectors, and the calculation process of the matrix elements is as follows:

[0106]

[0107] in, and The first one extracted from the current batch of training data The and the first 384-dimensional feature vector samples;

[0108] Steps 3-5 calculate the fractional-order decoupling penalty error based on the feature correlation matrix. The calculation process is as follows:

[0109]

[0110] in The total number of samples in the training data. Represents the trace operation of a matrix. The exponent of the fractional homomorphism and It is used to perform fractional-order damping attenuation of the coupling degree of nonlinear features in high-dimensional space. The centering matrix is ​​used to map features to a high-dimensional kernel projection space with zero mean. The calculation process of this centering matrix is ​​as follows:

[0111]

[0112] in, for The identity matrix, For length is A column vector of all 1s. In this embodiment, it is assumed that the total number of samples in the current batch of training data is... Construct a centralized matrix for mapping a high-dimensional kernel space with zero mean. as follows:

[0113]

[0114] Then, the fractional homomorphism exponent is set. Substitute the values ​​into the matrix above to calculate the fractional-order decoupling penalty error. The specific calculation process for the decoupling penalty error in this round is as follows:

[0115]

[0116] Among them, the decoupling penalty error calculated in this round Approximately 0.0097;

[0117] Steps 3-6 will Input is fed into the anisotropic information funnel module to predict the high-dimensional feature sphere. The direction mean vector on Concentration parameters and covariance deformation tensor The posterior probability density function of the feature is calculated using the prediction parameters. Its functional expression is as follows:

[0118]

[0119] in, For feature dimension, This is a thermodynamic distribution temperature scaling factor used to smooth the probability density surface of the manifold space in the early stages of training to prevent local gradient collapse. This is a high-dimensional volume constraint operator used to analyze the nonlinear mapping relationship between parameters and high-dimensional spatial volume, ensuring that the constructed posterior distribution satisfies the geometric constraint that the probability integral is always 1. In this embodiment, Inputting anisotropic information funnel module, the hyperspherical distribution parameters are predicted: direction mean vector. (dimension is) ), concentration parameters and covariance deformation tensor Combining feature dimensions With thermodynamic distribution temperature scaling factor Smooth surfaces, substitute with high-dimensional volume constraint operators Calculate the posterior probability density function of the feature. The function expression, after substituting the numerical values, is as follows:

[0120]

[0121] Assuming that after substituting the parameters, the expected value of the scalar measure of this probability density function distribution is approximately ;

[0122] Step 3-7 Calculate the probability density function With the target prior uniform distribution Feature compression regularization loss The loss is the distribution divergence between the two, and the calculation process is as follows:

[0123]

[0124] in, For high-dimensional characteristic spheres The prior uniform probability density on, The dynamic information capacity control threshold is used to preserve the basic semantic representation capacity during training and prevent semantic collapse caused by excessive compression of the feature manifold. In this embodiment, it combines the calculated scalar measure expectation value with the set dynamic information capacity control threshold. Calculate the posterior distribution and the target prior uniform distribution. Feature compression regularization loss The calculation process is as follows:

[0125]

[0126] Among them, the calculated feature compression regularization loss Approximately 0.4086;

[0127] Steps 3-8 combined , and Generalization features are obtained by performing spherical constrained noisy sampling. The sampling process is as follows:

[0128]

[0129] in It is a random noise matrix. This is a manifold curvature adaptive perturbation damping system used to dynamically adjust the injection intensity of noise based on the local curvature of a high-dimensional characteristic sphere. This is a spatial distance normalization operation. It uses Euclidean distance to force a strict projection and constraint of the features after adding local noise perturbation onto a given surface. On the manifold surface of the dimensional feature sphere, thus ensuring that the generalization feature satisfies the geometric prerequisites of the probability density function, in this embodiment, adaptive perturbation damping of the manifold curvature is combined. With the above concentration Substitute the prediction parameters into the calculation, and perform spherical constraint noise-added sampling. The calculation process is as follows:

[0130]

[0131] Through Euclidean distance ( The normalization constraint is forced onto the manifold, and the output is a 384-dimensional generalized feature containing local perturbations. (Due to the injection of noise, the first three elements of its tensor undergo a slight shift, changing from the original core feature values ​​to...) );

[0132] Steps 3-9 will generalize features Input evidence reasoning layer, output ambiguous words In each semantic category The amount of non-negative evidence below The calculation process is as follows:

[0133]

[0134] in, and These are the Laplace evidence reasoning layers corresponding to the first... The weight vectors and biases for each semantic category are used in this embodiment to generalize the features. Input a Laplace evidence inference layer, and use a fully connected layer with a Softplus activation function to output the non-negative evidence quantity of the ambiguous word "white spot" in three semantic categories. Assuming that after forward propagation calculation using the above formula, the amount of non-negative evidence output by the model in the current round is as follows: , , ;

[0135] Steps 3-10: Based on the amount of non-negative evidence Construct evidence distribution parameters and total amount of evidence And calculate the expected classification error. The calculation process is as follows:

[0136]

[0137]

[0138]

[0139] in, This is a long-tail category sensitivity compensation coefficient, used to adaptively amplify the weight of the medical rare disease category in gradient backpropagation. For topological focusing index, This is a logarithmic derivative operator used to calculate the logarithmic expectation of the polynomial prediction probability under the current evidence parameters. It smoothly maps the discrete classification error to a continuous distribution space, measuring the expected error of the model in the cognitive blind spot. For the one-hot encoding of the real label, in this embodiment, the real label of the sample is known to be... (Right now ), calculate the evidence parameters and total evidence amount for each semantic category according to the formula. The calculation process is as follows:

[0140]

[0141]

[0142]

[0143]

[0144] Subsequently, the long-tail compensation coefficient was set. Topology Focusing Index Calculate the expected classification error The calculation process is as follows:

[0145]

[0146] Among them, the expected classification error obtained in this round It is approximately 14.8683;

[0147] Step 3-11 Based on distribution parameters Calculate the logarithmic divergence penalty used to punish overconfidence. The calculation process is as follows:

[0148]

[0149] in, This is a continuous factorial extension operator used to analytically compute the multivariate integral normalization constant of the evidence distribution. It measures the global probability difference between the model's predicted distribution and the prior uniform distribution in the divergence integral term. It is a nonlinear truncation operator. To provide cognitive safety divergence margin, a reasonable feature ambiguity is preserved in the penalty process to prevent the model from losing its basic diagnostic intuition for similar clinical symptoms due to over-regularization. The formula above... The calculation process is as follows:

[0150]

[0151] in, In this embodiment, a cognitive safety divergence margin is set as a mask correction parameter. To prevent incorrect annealing of the correct category, it is first based on the true label. Calculate mask correction parameters Substitute the previous ones The numerical values ​​are:

[0152]

[0153]

[0154]

[0155] Total number of known semantic categories And the sum Substitute into the continuous factorial extension operator With the logarithmic derivative operator Calculate the logarithmic divergence penalty The calculation process is as follows:

[0156]

[0157] Among them, the logarithmic divergence penalty obtained in this round It is approximately 0.8420;

[0158] Step 3-12 will , , and The total functional error is obtained by weighted summation. The calculation process is as follows:

[0159]

[0160] in, , and The hyperparameter weights are based on the total functional error. Backpropagation is performed, updating the network parameters layer by layer until an optimized autocatalytic cognitive disambiguation network is obtained. In this embodiment, hyperparameter weights are set. , , The total functional error is calculated by weighted summation of the various errors. The calculation process is as follows:

[0161]

[0162] Among them, the calculated total functional error The error is approximately 14.9530. Subsequently, the model will propagate back along the original route based on this error and update the network parameters layer by layer to complete this round of optimization.

[0163] Step 4: The testing process, namely semantic classification and state collapse determination, involves inputting test data into the optimized autocatalytic cognitive disambiguation network and calculating ambiguous words. The confidence distribution and nonlinear cognitive state uncertainty under each semantic category are used to trigger rejection decisions through an uncertainty threshold. Among them, for safe samples that are not rejected, the semantic category with the highest confidence is the ambiguous word. The final semantic category, in this embodiment, is determined by performing semantic classification of the safety samples during the testing phase (i.e., scenario A). The medical record belonging to the test set, "The patient's oral mucosa shows [TGT]leukoplakia[ / TGT], accompanied by mild pain and ulcers," is fed into the optimized network for inference. The specific process is as follows:

[0164] Step 4-1: Input the test data into the optimized autocatalytic cognitive disambiguation network, which then passes through the BERT feature extractor and the feature manifold splitting network to extract the semantic backbone manifold features. In this embodiment, the test data containing the target marker is input into the network, and the pure semantic backbone manifold features of the test sample are extracted through forward propagation. ;

[0165] Step 4-2 will Input anisotropic Riemannian information funnel module to predict the direction mean vector During the testing phase, reparameterization sampling is stopped, and the direction mean vector is directly applied. As a generalization feature input to the Laplacian evidence inference layer, in this embodiment, the noisy sampling operation is stopped during the testing phase, and the direction mean vector predicted by the module is directly extracted. And send it as a deterministic feature into the inference layer.

[0166] Step 4-3 Extract the amount of non-negative evidence output from the Laplace evidence reasoning layer. Calculate the evidence distribution parameters and total amount of evidence Then calculate ambiguous words. In semantic categories Confidence level The calculation process is as follows:

[0167]

[0168]

[0169]

[0170] in The total number of semantic categories is known in this embodiment. The inference layer outputs explicit non-negative evidence for the test samples as follows: , , Based on the formula, the evidence parameters and total amount of evidence are reconstructed. The calculation process is as follows:

[0171]

[0172]

[0173]

[0174]

[0175] Then, the confidence level of the ambiguous term "white spot" under each category was calculated. The calculation process is as follows:

[0176]

[0177]

[0178]

[0179] This leads to the confidence distribution of the sample across various semantic categories;

[0180] Step 4-4 Based on parameters and Calculate the uncertainty of nonlinear cognitive state The calculation process is as follows:

[0181]

[0182] in, The state collapse sensitivity index and This is used to amplify cognitive uncertainty under high information entropy conditions in the feature space of extremely difficult medical records, thereby accelerating the triggering of the state collapse rejection mechanism in high-risk clinical scenarios. In this embodiment, a state collapse sensitivity index is set. Substituting into the formula for the uncertainty of nonlinear cognitive state, the calculation process is as follows:

[0183]

[0184] The nonlinear cognitive state uncertainty calculated in this study Approximately 0.1682;

[0185] Steps 4-5: Construct the state determination condition function To reduce the uncertainty of nonlinear cognitive states With the preset safety barrier threshold The comparison and segmented output of the judgment results are performed. The calculation process of the judgment condition function is as follows:

[0186]

[0187] Among them, when At that time, the current test data is judged to be an extremely difficult sample and is directly pushed into the manual review queue. When the current test data is determined to be a safe sample, the output with the highest confidence level is given. semantic categories As an ambiguous word The final predicted category;

[0188] In this embodiment, a preset safety barrier threshold is substituted. With the above ,because The system determines that the prediction is safe and outputs the value corresponding to the highest confidence level. (Oral leukoplakia) is the final predicted category;

[0189] Meanwhile, as a comparative calculation example (i.e., scenario B), if the input is a test medical record with extremely poor context: "Today's follow-up examination, [TGT]Vitiligo[ / TGT] has not shown significant shrinkage", the network will output a very low amount of evidence because it cannot extract effective modification information. , , This leads to the total amount of evidence obtained from the reconstruction. It is only 3.3737, at this time because The exponential amplification effect caused uncertainty to surge dramatically. If strict barrier thresholds are set for high-risk clinical scenarios at this time... ,because The system will proactively trigger the state collapse rejection mechanism based on the judgment condition function, refusing to output any blind guess results and pushing them into the manual review queue.

[0190] The medical semantic disambiguation method based on nuclear homomorphic decoupling and Riemann funnel in this invention can select highly pure semantic features and thoroughly solve the problems of small sample size and model overconfidence through hyperspherical constraint distribution and evidence reasoning. In particular, the introduction of the state collapse mechanism has extremely high clinical application safety and practical value.

[0191] The above is a detailed description of the embodiments of the present invention in conjunction with the accompanying drawings. The specific embodiments described herein are only for the purpose of helping to understand the method of the present invention. For those skilled in the art, based on the ideas of the present invention, changes and modifications can be made to the specific embodiments and application scope. Therefore, this specification should not be construed as a limitation of the present invention.

Claims

1. A medical semantic disambiguation method based on nuclear homomorphic decoupling and Riemannian funnel, for ambiguous words. have semantic categories Its characteristics are, The method includes the following steps: Step 1: Perform lexical segmentation, part-of-speech tagging, and semantic category tagging on the entire biomedical WSD corpus, obtain the part-of-speech tag sequence and semantic category sequence corresponding to each word, and locate ambiguous words. Target markers are inserted before and after the text to form a sequence of marked texts with an attention-guided mechanism. The processed corpus is then divided into training corpus and test corpus. Step 2: For the labeled text sequences, part-of-speech tag sequences, and semantic category sequences extracted from the training corpus of the biomedical WSD corpus, use the word segmenter of the pre-trained language model BERT to perform dictionary index mapping and convert them into input tensor sequences as training data. For the labeled text sequences, part-of-speech tag sequences, and semantic category sequences extracted from the test corpus of the biomedical WSD corpus, use the word segmenter of the pre-trained language model BERT to perform dictionary index mapping and convert them into input tensor sequences as test data. Step 3: Optimize the autocatalytic cognitive disambiguation network using training data to obtain the optimized autocatalytic cognitive disambiguation network. The autocatalytic cognitive disambiguation network is a model based on fractional-order kernel homomorphic decoupling and anisotropic Riemannian information funnel, which includes a BERT feature extractor, a feature manifold splitting network, a fractional-order regenerating kernel homomorphic decoupling module, an anisotropic Riemannian information funnel module, and a Laplace evidence reasoning layer. Step 4: The testing process, namely semantic classification and state collapse determination, involves inputting test data into the optimized autocatalytic cognitive disambiguation network and calculating ambiguous words. The confidence distribution and nonlinear cognitive state uncertainty under each semantic category are used to trigger rejection decisions through an uncertainty threshold. Among them, for safe samples that are not rejected, the semantic category with the highest confidence is the ambiguous word. The final semantic category.

2. The medical semantic disambiguation method based on nuclear homomorphic decoupling and Riemannian funnel according to claim 1, characterized in that, In step 1, the specific steps for preprocessing and segmenting the entire biomedical WSD corpus are as follows: Step 1-1: Use the text processing module to read Chinese sentences and corresponding ambiguous words from the biomedical corpus. The medical professional lexicon and natural language processing toolkit were used to perform word segmentation, part-of-speech tagging and semantic category tagging on the Chinese sentences to obtain a part-of-speech tag sequence and semantic category sequence corresponding to each word. Steps 1-2: Locating ambiguous words in Chinese sentences And insert special markers [TGT] and [ / TGT] before and after it respectively to form a marked text sequence with an attention guidance mechanism; Steps 1-3 divide the entire corpus, which includes the above-mentioned marked text sequences, part-of-speech tag sequences, and semantic category sequences, into training corpus and test corpus.

3. The medical semantic disambiguation method based on nuclear homomorphic decoupling and Riemannian funnel according to claim 1, characterized in that, In step 2, the construction of training and testing data involves the following steps: Step 2-1: For the divided training and test corpora, use the Tokenizer of the pre-trained language model BERT to perform word segmentation and dictionary index mapping on the marked text sequences to obtain discrete Token ID sequences. Step 2-2 involves uniformly truncating and padding the Token ID sequence to generate the input sequence tensor input_ids and the corresponding attention mask tensor attention_mask; Steps 2-3 extract the input_ids and attention_mask tensors corresponding to the training corpus, align them with their true semantic category labels, and convert them into PyTorch Dataset format; Steps 2-4 use DataLoader to encapsulate the training Dataset into iterable training data; Steps 2-5 extract the input_ids and attention_mask tensors corresponding to the test corpus and convert them into PyTorch Dataset format; Steps 2-6 use DataLoader to encapsulate the test Dataset into iterable test data.

4. The medical semantic disambiguation method based on nuclear homomorphic decoupling and Riemannian funnel according to claim 1, characterized in that, In step 3, the autocatalytic cognitive disambiguation network is trained and optimized. The specific steps are as follows: Step 3-1: Input the training data into the initialized autocatalytic cognitive disambiguation network and use the feature extractor to extract the pooling layer output features of the text. Step 3-2 inputs the corresponding part-of-speech tag sequence and semantic category sequence into independent embedding layers to map and obtain part-of-speech feature vectors and semantic category feature vectors. Then, the pooling layer output features, part-of-speech feature vectors, and semantic category feature vectors are multi-source heterogeneous concatenation and fusion, and the fused features are used as the initial homeomorphic manifold features. ; Step 3-3 will define the initial homeomorphic manifold features. The input is fed into a feature manifold splitting network, where parallel fully connected layers combined with the GELU activation function are used to map the features into semantic backbone manifold features. and background fluctuation manifold characteristics The mapping process is as follows: in, , This is the weight matrix. , It is the bias vector; Steps 3-4 will and The input is fed into the fractional-order regeneration kernel homomorphic decoupling module, which constructs feature correlation matrices based on the exponential distance between feature vectors. and The calculation process for the matrix elements is as follows: in, For bandwidth parameters, and The first one extracted from the current training data is the second one. The and the first 1 feature vector sample; Steps 3-5 calculate the fractional-order decoupling penalty error based on the feature correlation matrix. The calculation process is as follows: in The total number of samples in the training data. Represents the trace operation of a matrix. The exponent of the fractional homomorphism and It is used to perform fractional-order damping attenuation of the coupling degree of nonlinear features in high-dimensional space. The centering matrix is ​​used to map features to a high-dimensional kernel projection space with zero mean. The calculation process of this centering matrix is ​​as follows: in, for The identity matrix, For length is A column vector of all 1s; Steps 3-6 will Input is fed into the anisotropic information funnel module to predict the high-dimensional feature sphere. The direction mean vector on Concentration parameters and covariance deformation tensor The posterior probability density function of the feature is calculated using the prediction parameters. Its functional expression is as follows: in, For feature dimension, This is a thermodynamic distribution temperature scaling factor used to smooth the probability density surface of the manifold space in the early stages of training to prevent local gradient collapse. It is a high-dimensional volume constraint operator used to analyze the nonlinear mapping relationship between parameters and high-dimensional spatial volume, ensuring that the constructed posterior distribution satisfies the geometric constraint that the probability integral is always 1; Step 3-7 Calculate the probability density function With the target prior uniform distribution Feature compression regularization loss The loss is the distribution divergence between the two, and the calculation process is as follows: in, For high-dimensional characteristic spheres The prior uniform probability density on, The dynamic information capacity control threshold is used to preserve the basic semantic representation capacity during training and prevent semantic collapse caused by excessive compression of the feature manifold. Steps 3-8 combined , and Generalization features are obtained by performing spherical constrained noisy sampling. The sampling process is as follows: in It is a random noise matrix. This is a manifold curvature adaptive perturbation damping system used to dynamically adjust the injection intensity of noise based on the local curvature of a high-dimensional characteristic sphere. This is a spatial distance normalization operation. It uses Euclidean distance to force a strict projection and constraint of the features after adding local noise perturbation onto a given surface. On the manifold surface of the 3D feature sphere, thereby ensuring that the generalization feature satisfies the geometric prerequisites of the probability density function; Steps 3-9 will generalize features Input evidence reasoning layer, output ambiguous words In each semantic category The amount of non-negative evidence below The calculation process is as follows: in, and These are the Laplace evidence reasoning layers corresponding to the first... Weight vectors and biases for each semantic category; Steps 3-10: Based on the amount of non-negative evidence Construct evidence distribution parameters and total amount of evidence And calculate the expected classification error. The calculation process is as follows: in, This is a long-tail category sensitivity compensation coefficient, used to adaptively amplify the weight of the medical rare disease category in gradient backpropagation. For topological focusing index, This is a logarithmic derivative operator used to calculate the logarithmic expectation of the polynomial prediction probability under the current evidence parameters. It smoothly maps the discrete classification error to a continuous distribution space, measuring the expected error of the model in the cognitive blind spot. One-hot encoding for the real label; Step 3-11 Based on distribution parameters Calculate the logarithmic divergence penalty used to punish overconfidence. The calculation process is as follows: in, This is a continuous factorial extension operator used to analytically compute the multivariate integral normalization constant of the evidence distribution. It measures the global probability difference between the model's predicted distribution and the prior uniform distribution in the divergence integral term. It is a nonlinear truncation operator. To provide cognitive safety divergence margin, a reasonable feature ambiguity is preserved in the penalty process to prevent the model from losing its basic diagnostic intuition for similar clinical symptoms due to over-regularization. The formula above... The calculation process is as follows: in, Correct parameters for the mask; Step 3-12 will , , and The total functional error is obtained by weighted summation. The calculation process is as follows: in , and The hyperparameter weights are based on the total functional error. Backpropagation is performed to update the network parameters layer by layer until the optimized autocatalytic cognitive disambiguation network is obtained.

5. The medical semantic disambiguation method based on nuclear homomorphic decoupling and Riemannian funnel according to claim 1, characterized in that, In step 4, reasoning and state collapse determination are performed on the test data. The specific process is as follows: Step 4-1: Input the test data into the optimized autocatalytic cognitive disambiguation network, which then passes through the BERT feature extractor and the feature manifold splitting network to extract the semantic backbone manifold features. ; Step 4-2 will Input anisotropic Riemannian information funnel module to predict the direction mean vector During the testing phase, reparameterization sampling is stopped, and the direction mean vector is directly applied. As a generalization feature, it is input into the Laplace evidence reasoning layer; Step 4-3 Extract the amount of non-negative evidence output from the Laplace evidence reasoning layer. Calculate the evidence distribution parameters and total amount of evidence Then calculate ambiguous words. In semantic categories Confidence level The calculation process is as follows: in The total number of semantic categories; Step 4-4 Based on parameters and Calculate the uncertainty of nonlinear cognitive state The calculation process is as follows: in, The state collapse sensitivity index and It is used to steeply amplify the cognitive uncertainty in the high information entropy state in the feature space of extremely difficult medical records, thereby accelerating the triggering of the state collapse rejection mechanism in high-risk clinical scenarios. Steps 4-5: Construct the state determination condition function To reduce the uncertainty of nonlinear cognitive states With the preset safety barrier threshold The comparison and segmented output of the judgment results are performed. The calculation process of the judgment condition function is as follows: Among them, when At that time, the current test data is judged to be an extremely difficult sample and is directly pushed into the manual review queue. When the current test data is determined to be a safe sample, the output with the highest confidence level is given. semantic categories As an ambiguous word The final predicted category.

Images