A medical terminology standard concept mapping method and system

Abstract

This invention discloses a method and system for mapping standard concepts in medical terminology. It retrieves a candidate set of target mentions from medical texts, calculates the candidate set coverage probability based solely on the recall coverage features before feedback from closed-set relations, and obtains the candidate conditional probability by calibrating the candidate scores within the set. The candidate conditional probability before freezing and updating is captured as a snapshot, used to determine competing candidates and distinguishing terms, and to request enumeration relations from a generative language model. The complete relations are centered according to the normalized probability of this snapshot to obtain a weighted zero-sum candidate relative feedback. Within the item-by-item boundary and the total budget, the shared term weights and sign-preserving candidate residuals are reconstructed from the initial values. The candidate conditional probabilities are then updated through monotonic scoring and shared calibration, and standard concept identifiers are output. This isolates the uncertainty of in-set ranking and candidate coverage, and makes the model output verifiable bounded evidence.

Description

Technical Field

[0001] This invention relates to the fields of natural language processing and medical information standardization, and particularly to a method and system for mapping standard concepts in medical terminology. This invention maps entity references in electronic medical documents, imaging reports, test descriptions, and follow-up records to standard concept identifiers in a standard terminology database. Accepted standard concept identifiers are then written into structured medical data records for retrieval, statistics, and master data governance, without outputting status judgment conclusions. Background Technology

[0002] Medical texts are primarily written freely, and the same standard concept often appears in different source documents with full names, abbreviations, alternative names, and context-dependent spellings. Furthermore, different standard concepts share a large number of location terms, grade terms, and degree terms. To include these texts in the statistical and retrieval process, entity references in the text must first be mapped to unique standard concept identifiers in a standard terminology database. The accuracy and reversibility of this mapping directly determine the quality of downstream data governance.

[0003] Existing mapping schemes can be broadly categorized into three types. The first type pre-assigns fixed weights to location words, state words, and degree words based on word frequency, inverse document frequency, or a human-generated dictionary, and then calculates the literal similarity between the text to be mapped and the candidate standard concepts. The second type matches text vectors with concept vectors, or uses multi-channel recall combined with a deep ranking model for fine-grained ranking. The third type allows generative language models to directly participate in the generation, selection, or re-ranking of candidates, or uses generative language models to expand the training corpus before recall and ranking.

[0004] The closest approach to this invention is the third type, where the model outputs matching labels and probabilities for each candidate pair. This approach presents the text to be mapped and all candidates to the model, which then provides a label indicating whether each candidate matches and its self-reported probability. The model verifies the match by swapping the order in which the candidates are presented, and ultimately selects one candidate or outputs "none of the above" in the same decision. Therefore, the model's output simultaneously performs the functions of evidence generation and result adjudication; both the ranking result and the candidate coverage judgment are given in the same model call.

[0005] The above approach has two limitations. First, the model's output is directly used as the mapping result or directly superimposed on the candidate scores. The model may return identifiers outside the given candidate set, is sensitive to the order of candidate presentation, and its self-reported probabilities are not calibrated by external annotations. Due to the lack of verifiable formal constraints in the intermediate process, any missing, overweighted, or delayed responses will result in errors being directly written into the mapping result, and these errors are difficult to reproduce or revert because the parameters are rewritten online. Second, the support components contributed by terms shared by multiple competing candidates will simultaneously inflate the scores of all competing candidates, making it difficult to locate the original terms that truly differentiate. At the same time, a sharp probability distribution within the candidate set is often taken as evidence of sufficient recall. When the correct concept is not recalled at all, the solution will still select the first candidate from the existing candidates and submit it, resulting in an inconspicuous erroneous submission. Summary of the Invention

[0006] The technical problems to be solved by this invention are as follows: First, when the output of a generative language model is directly used as a mapping result or directly superimposed on the candidate score, there are problems such as exceeding the candidate set, sensitivity to the order of candidate presentation, self-reported confidence not being externally calibrated, and missing items and overweighted responses being directly written into the mapping result. Moreover, online parameter rewriting makes the results difficult to reproduce and roll back. Second, terms shared by multiple competing candidates cause the common support components to simultaneously raise the scores of each competing candidate, making it difficult to locate the terms that truly produce a distinguishing effect. At the same time, the sharpening of the distribution within the candidate set is mistakenly taken as sufficient evidence of recall, and erroneous submissions will still occur when the correct concept is not recalled.

[0007] To address the aforementioned technical problems, this invention provides a method for mapping standard concepts in medical terminology, comprising:

[0008] Construct mapping units based on target references in medical texts, and recall candidate sets from a standard terminology database;

[0009] The term-weighted sum of shared term weights and initial zero candidate residuals is used as the candidate directed feedback feature. The candidate score is obtained by a fixed scoring function and calibrated as the candidate conditional probability. The candidate set coverage probability is calculated only using the recall coverage feature before feedback of closed set relation.

[0010] The candidate conditional probabilities before the update are frozen as a snapshot, based on which the competing candidate set and distinguishing terms are determined. After a closed set identification request, the generative language model returns only the enumeration relationship of the distinguishing terms to the competing candidate set, and after integrity verification, it is converted into bounded signed relation values.

[0011] The complete relation is centered according to the normalized probability of the snapshot in the competing candidate set to obtain the candidate relative feedback with a weighted sum of zero, based on which the shared term weights and sign-preserving candidate residuals are reconstructed from the initial values ​​within the item-by-item boundary and the total budget.

[0012] By re-scoring the candidate directional feedback features using the fixed scoring function that does not decrease monotonically, updating the candidate conditional probabilities through shared monotonic calibration, and outputting the standard conceptual identifier of the first candidate, which is then written into the medical data record.

[0013] Furthermore, freezing the candidate conditional probabilities before the update into a snapshot to determine the competing candidate set and distinguishing terms includes discretizing the basic matching strength into signature bins according to a registered threshold, calculating the signature information gain of each term using the normalized probability of the snapshot in the competing candidate set and the signature bins, determining the query value of each term through a pre-registered combination function, selecting up to a preset number of terms in descending order of the query value to form the distinguishing terms, and not sending the closed set identification request when the query values ​​are all lower than the registered value threshold or there are fewer than two competing candidates.

[0014] Furthermore, the signature information gain is determined as the difference between the entropy of the candidate random variable and the conditional entropy of the candidate random variable under the conditions of the signature random variable. Signature bins with a probability mass of zero do not participate in the summation of the conditional entropy. When all competing candidate signature bins are the same, the signature information gain is zero, and negative values ​​caused by numerical errors are clipped to zero.

[0015] Furthermore, the closed set identification request and the response to the request are simultaneously bound to an order-independent candidate set summary, an order-dependent candidate request summary, and a relational domain identifier. If the response is missing any of the requested term-candidate pairs, the entire request is deemed invalid. If the equivalent requests with different candidate presentation orders return inconsistent results, the corresponding term-candidate pairs are deemed unobserved.

[0016] Furthermore, the bounded signed relation value is obtained by converting the enumerated relation into a signed expectation using the conditional confusion matrix of the relation calibration group. The relation calibration group is enabled when the sample size and the lower bound of the confidence interval of the performance index reach the registration threshold; otherwise, it is rolled back or disabled according to a preset level, and the confidence level reported by the model is not used.

[0017] Furthermore, the method also includes statistically analyzing common exclusion indicators and dispersed exclusion indicators before centralization and reconstruction. When any indicator reaches the corresponding threshold, a coverage alarm is triggered. All parameters and two-level probabilities are kept at their previous values, and the competitive candidate set is expanded or supplementary recall is performed.

[0018] Furthermore, the reliability of the relationship is the minimum value among the relationship calibration reliability factor, the source reliability factor, and the structural verification reliability factor. Each factor value is between zero and one and does not depend on the sign of the relationship value.

[0019] Furthermore, the centralization is performed only on complete relation rows, and relation rows with unobserved terms and candidate pairs are set to zero as a whole. In each feedback round, all complete relation rows accumulated in the current relation domain are uniformly re-centralized using the snapshot of this round.

[0020] Furthermore, the reconstruction includes determining the term differentiation strength using a weighted L2 norm, performing a multiplicative update of the initial shared term weights and projecting them onto the multiplier box and the L1 norm total budget feasible region, and projecting the candidate residual target value onto a feasible region that simultaneously satisfies the terms-by-term absolute value boundary, the L1 norm total budget, the weighted zero mean, and the sign-preserving constraint.

[0021] Furthermore, the candidate residual can be represented by a complete relational row as the product of the same nonnegative row coefficient and the candidate relative feedback of that row, wherein the nonnegative row coefficient is constrained by the term-by-term absolute value boundary and the first norm total budget.

[0022] Furthermore, the candidate set coverage probability and the candidate conditional probability are isolated from each other through data paths. When the candidate set summary remains unchanged, the candidate set coverage probability and the candidate set out-of-set probability are kept as read-only and are recalculated only when a new candidate generation is established.

[0023] Furthermore, the method also includes maintaining a pre-update checkpoint and a separate monotonic resource ledger, atomically committing the mapping data state when all assertions pass, otherwise rolling back the entire data, and the rollback does not restore the deducted resources.

[0024] This invention also provides a medical terminology standard concept mapping system, including a mapping unit construction module, an initial scoring and two-level probability module, a closed set identification and verification module, a centering and bounded reconstruction module, and a monotonic rearrangement and output module.

[0025] The technical effects of this invention are as follows: By limiting the function of the generative language model to returning relations on fixed term identifiers, fixed candidate identifiers, and fixed relation enumerations, and by binding requests and responses to the same summary and relation domain, and by deeming the entire request term-candidate pair invalid if any requested term is missing, and by converting the verified enumerated relations into bounded signed relation values ​​through an externally labeled conditional confusion matrix, the model output can only enter the scheme as intermediate evidence that can be verified item by item and is numerically bounded. Identifiers that exceed the candidate set, missing responses, and late responses are intercepted before entering parameter updates. Furthermore, since the shared term weights and candidate residuals are reconstructed from the initial values ​​in each round instead of being multiplied on the results of the previous round, and the reconstruction results are constrained by item-by-item boundaries and the total budget, and are rolled back to the checkpoint before the update when the assertion fails, the problems of model output being directly written into the mapping results and parameters being rewritten online and difficult to reproduce and rollback are solved. This allows for deterministic recalculation of the mapping process under the same input and the same version. Because the complete relation is centered using the normalized probability of the same pre-update candidate conditional probability snapshot within the competing candidate set, the common components of each competing candidate obtaining the same relation value are strictly canceled out after centering. Only the difference components that can change the relative relationship between candidates participate in the reordering. Furthermore, since the candidate set coverage probability is calculated only based on the recall coverage features formed before the feedback of the closed set relation, and remains read-only when the candidate set summary has not changed, a sharper probability distribution within the set cannot improve the candidate coverage probability. Moreover, the two indicators of common exclusion and dispersed exclusion can block the submission of this batch before the parameter update. This solves the problems of common terms raising all competing candidates and sharp distributions within the candidate set being mistakenly considered as sufficient recall, resulting in erroneous submissions. This allows the uncertainty of the in-set ranking and the uncertainty of candidate coverage to be measured and handled separately. Attached Figure Description

[0026] Figure 1 A flowchart illustrating the medical terminology standard concept mapping method provided in this embodiment of the invention;

[0027] Figure 2 A flowchart illustrating the steps of determining the competitive candidate set and distinguishing terms provided in an embodiment of the present invention;

[0028] Figure 3 A flowchart illustrating the coverage alarm determination and branch handling steps provided in an embodiment of the present invention;

[0029] Figure 4 This is a flowchart illustrating the shared term weight and candidate residual dual-channel bounded reconstruction steps provided in an embodiment of the present invention.

[0030] Figure 5 A schematic diagram of the module structure of the medical terminology standard concept mapping system provided in an embodiment of the present invention. Detailed Implementation

[0031] The present invention will be further described below with reference to embodiments. The parameters, standard concept names, and example probabilities mentioned are used to explain the technical process and do not constitute a limitation on the terminology system, model type, threshold, or deployment method.

[0032] Example 1

[0033] This embodiment provides a method for mapping standard concepts in medical terminology, such as... Figure 1 As shown, the method includes:

[0034] S1: Construct mapping units based on target references in medical texts and recall candidate sets from a standard terminology database.

[0035] The identification, text normalization, and multi-channel retrieval of the target mentions can be achieved using conventional methods in the field, such as dictionary matching, sequence labeling, character approximation retrieval, and vector retrieval; this invention does not limit these methods. The mapping unit uses a unique target mention as an anchor, and neighboring mentions within the same segment serve only as binding contexts verified by scope. Segments containing multiple targets to be mapped are split into multiple mapping units according to the target.

[0036] Specifically, the mapping unit is divided based on sentence boundaries, parallel conjunctions, transition words, colons, and explicit state boundaries. When multiple targets to be mapped appear in the same segment, they are split according to the target, so that the negation scope and degree modification do not spread across units. Each term obtained from word segmentation stores a normalized value, original span, whether it is located within any entity mention, part of speech or medical type, optional context state, unit type, and source view. The term identifier within the request is jointly determined by the request identifier, mapping unit identifier, and span, and is used for binding the closed set result. A stable term signature without the request identifier is generated separately for cache reuse within the authorized scope.

[0037] In this embodiment of the invention, the recall can be performed in parallel by one or more of the following channels: lexical channel, alias channel, abbreviation channel, character approximation channel, vector semantic channel, and structured constraint channel. Each channel independently returns candidate identifiers and rankings within the channel without overwriting the results of other channels. The reciprocal of the rankings within each channel is taken and weighted by the channel reliability coefficient to obtain the candidate's recall score. Candidates not recalled by a certain channel have zero contribution to that channel. A maximum of a preset number of candidates are truncated in descending order of recall score, and hard constraints such as valid identifier, valid concept, and compatible type are applied. It should be noted that attributes such as source department and examination modality are preferably used as calibrable features or hard constraints with clear evidence to avoid excluding correct concepts that actually exist simply because the source department is inconsistent.

[0038] It is worth noting that this step performs deterministic handling of several boundary cases: when the candidate set is empty and all recall channels are healthy, the coverage probability of the candidate set is set to zero and supplementary recall is performed if the budget allows; if the budget is insufficient, the status outside the candidate set is output; when a channel fails, a degraded status is output instead of representing infrastructure failure as unmatched; when the number of candidates is one, the candidate conditional probability of that candidate is set to one and the joint probability of the second candidate is set to zero, but it still needs to pass conditions such as coverage probability, target boundary and structural conflict before it can be submitted; when the number of terms is zero, it directly enters the candidate set outside branch or the manual review branch.

[0039] S2: The term-weighted sum of shared term weights and initial zero candidate residuals is used as the candidate directed feedback feature. The candidate score is obtained by a fixed scoring function and calibrated as the candidate conditional probability. The candidate set coverage probability is calculated only using the recall coverage feature before the closed set relation feedback.

[0040] Specifically, this step consists of five stages. First, the mapping unit and its bound context are segmented to obtain a sequence of terms. The original importance of each term is calculated based on inverse document frequency, medical dictionary importance, whether the term is located within an entity, and whether it has a negative or degree modification effect. And normalized to the initial shared term weights, as shown in Equation (1):

[0041] Equation (1):

[0042] in, Indicates terms, This represents the set of terms for the mapping unit. Indicates terms The original importance, Indicates terms The initial shared term weights, and Secondly, a signature is pre-established for each candidate standard name, alias, character fragment, phonetic encoding, and structured tag in the standard terminology database, and the term is calculated. With candidates Basic matching strength between Its value is pruned to between zero and one and remains fixed within the same candidate generation; candidate The text coverage is shown in equation (2):

[0043] Equation (2):

[0044] in, Indicates terms The current shared term weights, in the initial state , Indicates candidate The text coverage rate. Third, let the candidate residuals The initial value is zero, and the candidate directional feedback feature is defined as the sum of the product of the shared term weight and the candidate residual, as shown in equation (3):

[0045] Equation (3):

[0046] in, Indicates terms For candidates candidate residuals Indicates candidate Candidate directional feedback features, in the initial state Fourth, the candidate basic feature vectors and The fixed scoring function is input together, as shown in equation (4):

[0047] Equation (4): In the linear implementation

[0048] in, Indicates candidate The basic feature vector includes the text coverage, candidate coverage, character similarity, alias consistency, vector semantic similarity, and structural compatibility. This represents a fixed scoring function with constant online parameters. Represents the basic feature coefficient vector. The coefficients of the candidate directional feedback features and , Indicates candidate The candidate scores are then used. Fifth, the candidate set coverage probability is obtained using the recall coverage features via a coverage calibrator. The candidate conditional probability is obtained by calibrating the candidate scores within the set through shared criteria. The joint distribution is shown in equation (5):

[0049] Equation (5): ,

[0050] in, This represents the candidate set coverage probability that the correct concept is located within the candidate set. Indicates candidate Candidate conditional probability and , Indicates candidate The joint probability, This represents the probability of being outside the candidate set.

[0051] Its technical significance lies in the isolation of the two probabilistic pathways. The reason why the candidate set coverage probability is required to be calculated solely from the recall coverage features formed before the feedback of the closed-set relation is that the feedback of the closed-set relation can only change the relative merits between candidates, and does not carry information about whether the correct concept is not even in the set. If the feedback quantity is allowed to enter the input of the coverage probability, the coverage probability will increase when the probability distribution within the set becomes sharp due to the feedback, causing the failure mode of recall omission to be masked by its own ranking confidence. Since the input of the coverage probability is limited to the deterministic features of the recall stage, any sharpening of the distribution within the set cannot increase the coverage probability, thus allowing erroneous submissions when the correct concept is not recalled to be identified at the probabilistic level, alleviating the problem in the background technique where a sharp distribution within the set is mistakenly taken as sufficient recall. Regarding parameters, The score must be strictly greater than zero to ensure that the score monotonically increases with respect to candidate directional feedback features. When using a nonlinear ranking model, shape constraints or pre-launch monotonicity checks should be used instead. The coverage calibrator must be trained on an independent calibration set containing artificially constructed candidate omission negative samples; otherwise, its output will be systematically overestimated. In one example, the target mention recall of an image examination report text yields two candidate concepts from a standard terminology corpus, and the coverage calibrator outputs... The conditional probability within the set is , The remaining candidates total ,but , , The joint probability interval between the first and second positions is insufficient for direct submission.

[0052] Furthermore, the candidate coverage rate in the candidate basic feature vector and the text coverage rate constitute a bidirectional coverage metric: the text coverage rate measures the proportion of term weights matched by the candidate signature in the mapping unit, while the candidate coverage rate measures the proportion of components covered by the input text in the candidate's standard name or individual alias. The latter is calculated separately for the standard name and each alias, and the maximum value is taken to avoid increasing the denominator due to an increase in the number of aliases. The introduction of bidirectional coverage suppresses the situation where a high score is obtained only because of a few shared terms, while the key components of the candidate itself are not covered. The remaining components of the candidate basic feature vector include character or edit similarity, alias and abbreviation consistency features, vector semantic features, negation and degree and numerical compatibility features, and part and system and modality and entity type compatibility features; except for the text coverage rate component, which is recalculated with the reconstruction of the shared term weights, the remaining components remain fixed within the same candidate generation from the checkpoint reading.

[0053] In one implementation, a direct submission branch is provided after this step: when the joint probability of the first candidate is not lower than the direct submission threshold, the interval between the joint probabilities of the first and second candidates is not lower than the interval threshold, the out-of-set probability of the candidate is not higher than the out-of-set threshold, the entity boundary is valid and there are no conflicts in negation and type, and at least two independent signal families jointly support the first candidate, the generative language model can be directly submitted without invoking it. The division of the signal families, the transformation of the support scores, and their respective thresholds are all bound to versions. Signals derived from the same original text clue or the same basic data must not be counted repeatedly across families—standard name exact match, alias inclusion, and character overlap are only counted as one lexical family. Structured context, semantic retrieval, and closed-set relations through gating can constitute other signal families respectively. It should be understood that using at least two independent signal families is a highly reliable implementation method, and other verified thresholds can be used in lower-risk application scenarios.

[0054] The system also calculates normalized entropy and first-order interval based on the candidate condition probabilities for process monitoring. When probabilities are tied, a deterministic total order is formed by ascending order of immutable candidate identifiers. It should be noted that entropy only describes the degree of uncertainty, and a sharpening of the distribution within a set does not constitute a sufficient condition for a correct mapping. Therefore, the decrease in normalized entropy and the increase in candidate scores do not participate in the determination of mandatory acceptance conditions, so as to avoid substituting process quantities for acceptance quantities.

[0055] S3: Freeze the candidate conditional probabilities before the update into a snapshot, thereby determining the competing candidate set and distinguishing terms. Through the closed set identification request, the generative language model returns only the enumeration relationship of the distinguishing terms to the competing candidate set, and converts it into bounded signed relation values ​​through integrity verification.

[0056] Specifically, this step uses the same pre-update state in three places simultaneously, thus establishing word selection, requests, and subsequent centralization on the same benchmark. The snapshot freezes the candidate condition probabilities that have been accepted at the start of this round; the competitive candidate set is determined by the snapshot in descending order of condition probability and a score difference threshold, retaining at least two and at most a preset number of candidates. If the upper limit is exceeded, candidates are truncated using a fixed sorting key consisting of descending candidate condition probability, ascending score difference with the first candidate, and ascending immutable candidate identifier; the distinguishing terms are determined by the signature difference between the snapshot and the basic matching strength, and its specific calculation is as follows: Figure 2 As described in Embodiments 2 and 3, the closed set identification request consists of a request fact area, a candidate definition area, an allowed relation area, and an output format area. The allowed relation area lists only three enumerated values: support, exclusion, and irrelevance, corresponding to bounded signed codes. , and Generative language models must not create candidate identifiers, rewrite the original text span, or output final concept identifiers. The returned response undergoes format, request binding, membership, uniqueness, completeness, and overreach checks in sequence. The completeness check requires the response to cover all terms / candidate pairs of the distinguishing term and the competing candidate set; if any pair is missing, the entire response is considered invalid, not partially completed. The enumerated relations that pass all checks are then converted into bounded, signed relation values ​​through relation calibration. The conversion method is as described in Example 5.

[0057] The reason for limiting the model's output to enumerated values ​​on fixed term identifiers and fixed candidate identifiers, and handling responses with the principle of invalidation for missing terms, is that only outputs with both the value domain and identifier domain fixed in advance can be verified item by item by machine: a fixed identifier domain ensures that out-of-bounds candidates are rejected at the member check; a fixed value domain prevents responses from including conclusions in free text form; bidirectional binding of summaries causes late responses and cross-request responses to be discarded due to summary mismatch; and the requirement of dense integrity prevents the model from implicitly exercising its selection power by only answering a few pairs it is "confident" about. Due to the combined effect of these constraints, the model output is compressed into a verifiable, recalculated, and numerically bounded relation matrix before entering parameter updates, thus solving the problems of out-of-bounds model output, order sensitivity, and direct writing of missing responses into the mapping results in the background technology. Regarding parameters, a value that is too small for the number of competing candidates will miss the real competitors and cause the centralization to lose its reference; a value that is too large will cause the relation grid size and call overhead to grow linearly with the number of candidates. In practice, two to five is appropriate; the number of distinguishable terms is similarly constrained by the upper limit of registration.

[0058] S4: Center the complete relation according to the normalized probability of the snapshot in the competing candidate set to obtain the candidate relative feedback with a weighted sum of zero, thereby reconstructing the shared term weights and sign-preserving candidate residuals from the initial values ​​within the item-by-item boundary and the total budget.

[0059] Specifically, this step includes two parts: relativization and dual-channel reconstruction. The relativization part first determines the cumulative observation mask according to the term-candidate pair, and determines the term row that has been observed for all competing candidates as the complete relation row, and sets the rest of the relation rows to zero; then it normalizes the snapshot within the competing candidate set, as shown in Equation (6):

[0060] Equation (6): ,

[0061] in, This represents the set of competing candidates. Indicates candidate Normalized pre-update conditional probabilities within the competing candidate set. For each complete relation row... Centralization, as shown in equation (7):

[0062] Equation (7):

[0063] in, Indicates terms For candidates Bounded signed relation values, Indicates terms For candidates The relative feedback of the candidates, and always The reconstruction section first uses candidate relative feedback regarding... The weighted L2 norm determines the term distinguishing strength, as shown in Equation (8):

[0064] Equation (8):

[0065] in, Indicates terms The distinguishing strength. The shared term weight channel is projected after a multiplicative update from the initial value, as shown in equation (9):

[0066] Equation (9):

[0067] in, This represents the non-negative update step size of the shared term weight channel. Indicates to feasible region The projection operator, The construction is as described in Example 9. The candidate residual channel is projected from zero value into a bounded state, as shown in Equation (10):

[0068] Equation (10): ,

[0069] in, This represents the non-negative update step size of the candidate residual channel. Indicates terms The relation validity gating value of the row containing the relation and , Indicates the candidate residual target value. Indicates to feasible region The projection operator.

[0070] The causal chain in this step has two aspects. First, since the centralization uses the normalized probability of the same snapshot before the update as weights and only applies to complete relation rows, when the same term gives the same relation value to all competing candidates, the relative feedback of the candidate in that row is strictly zero. This prevents terms such as part words and hierarchical words, which are shared by multiple candidates, from conveying a relative advantage to any candidate. Only relations with different values ​​among candidates generate targeted updates, thus solving the problem in the background technology where the common support component of shared terms simultaneously raises the level of each competing candidate. Setting incomplete relation rows to zero as a whole, rather than filling them with default values, avoids unobserved terms being implicitly treated as "irrelevant" and diluting the centralization benchmark. Second, since both channels reconstruct from the initial shared term weights and zero-value candidate residuals rather than multiplying on the results of the previous round, and the reconstruction results are constrained within the item-by-item boundaries and the total change budget, the impact of any round of feedback on the parameters has a calculable upper bound, and cross-round errors do not accumulate round by round. This controls parameter drift under multi-round feedback and allows for deterministic recalculation in each round. Regarding parameters, and If the value is too small, the legitimate feedback will be difficult to change the candidate order and more feedback rounds and model calls will be required. If the value is too large, a single round of feedback can cross the boundary and frequently trigger pruning and rollback. In practice, the candidate order improvement rate and rollback rate should be used together to calibrate the independent development set.

[0071] In one example, following the example state of S2. , ; Partial words The bounded signed relation value is , Its weighted average is Therefore , Verification results Common hierarchical words between the two candidates Since the two candidates return the same relation value, their relative feedback is zero. And the boundary conditions were not triggered, so... , After reconstruction ,but , .

[0072] S5: By re-scoring the candidate directional feedback features using the fixed scoring function that does not decrease monotonically, updating the candidate conditional probability through shared monotonic calibration, and outputting the standard concept identifier of the first candidate into the medical data record.

[0073] Specifically, the re-scoring uses the same fixed scoring function and the same version parameters as in S2: the text coverage is recalculated according to Equation (2) using the reconstructed shared term weights; the candidate coverage, recall score, and structural compatibility, which are not updated by the feedback path authorization, continue to take the candidate generation checkpoint values, and then the updated candidate directional feedback features and candidate scores are obtained according to Equations (3) and (4). The candidate conditional probability is recalculated by a calibration transformation that is shared by all candidates, takes positive values, and strictly increases with respect to the candidate scores, as shown in Equation (11):

[0074] Equation (11):

[0075] in, The temperature parameter representing the calibration transformation and This transformation is equivalent to the candidate permutation, and therefore does not reverse the candidate score order within the same state. The candidate set coverage probability remains read-only when the candidate set summary remains unchanged, and its read-only condition and the re-estimation of the new candidate generation are as described in Example 11.

[0076] The reason for imposing the constraint that the scoring function is "monotonically non-decreasing for candidate-oriented feedback features" and requiring the calibration transformation to be shared and strictly monotonic across all candidates is that the correctness of the feedback path does not solely depend on the sign of the feedback value itself: if the scoring function has a decreasing range for that feature, a single legitimate feedback expressing "support" might actually lower the score of the corresponding candidate; if independent calibration mappings are allowed for different candidates, even if the score order is correct, the probability order might still be reversed by the candidate-by-candidate mapping. Because these two monotonicities are explicitly constrained and verified before deployment, "if the relation is supportive, the candidate is not degraded; if the relation is repulsive, the candidate is not ascended" becomes a structural property of the scheme rather than a training result, thus maintaining the direction of closed-set relation feedback throughout the entire chain. Continuing with the previous example, let's take... , The increment of the score difference between the two candidates is Logarithmic difference Become Therefore, after the update , The joint probability is , ,and Keep Unchanged – the feedback of the closed set relation redistributes the probability within the set, but does not change the candidate coverage probability.

[0077] Example 2

[0078] This embodiment further refines the step in Embodiment 1 of "freezing the candidate conditional probabilities before the update as a snapshot, thereby determining the competing candidate set and distinguishing terms." For example... Figure 2As shown, this feature includes:

[0079] S3.1: Discretize the basic matching strength of each candidate in the candidate set on each term into signature bins according to the registration threshold;

[0080] S3.2: Calculate the signature information gain of each term using the normalized probability of the snapshot in the competing candidate set and the signature binning;

[0081] S3.3: Using a pre-registered combination function, the query value of each term is determined by taking the signature information gain, the term being located within the target mention, the medical type of the term, the negation domain of the term, and the span quality of the term as gain terms, and the redundancy of the term relative to the selected terms and the transmission cost of the sensitive information of the term as subtraction terms.

[0082] S3.4: Select up to a preset number of terms in descending order of query value to form the distinguishing terms;

[0083] S3.5: When the query values ​​are all lower than the registered value threshold or there are fewer than two candidates in the competitive candidate set, the closed set identification request is not sent, and the process is switched to supplementary recall or manual review mark is output.

[0084] Signature binning uses two pre-registered thresholds. Deterministically discretize into three levels: no match, weak match, and strong match, denoted as . Since the binning threshold and the basic matching strength are both fixed within the same candidate generation, the binning results under the same state can be deterministically recalculated. The combination function of query value, the parameters of each factor, and the rules for handling parallel processing are all bound to the version and fixed before feedback begins.

[0085] The technical significance of this word selection method lies in transforming the question of "whether a term is important" into "whether a term can distinguish the current group of candidates." The static weighting scheme pre-configures fixed weights for location terms, state terms, and degree terms, and these weights are independent of the current competitive relationship. Since query value is primarily based on the signature distribution of the current competing candidates, the same term acquires different query values ​​in different competitive scenarios—hierarchical terms matched by all competing candidates have identical binning and zero information gain, while location terms matching only some candidates gain high gain. This ensures that the limited model call budget is allocated to terms that truly differentiate candidates. Sensitive information transmission costs are deducted, preventing content unrelated to the target mention and its binding context from being sent out due to its information gain. A preset number that is too small will prevent a single request from covering enough terms to differentiate candidates, requiring additional feedback rounds; a value that is too large will cause the relational grid size and load to increase linearly. In practice, a value of two to five, combined with redundancy suppression, is preferable.

[0086] Furthermore, in the multi-round closed-set relation feedback, each round recalculates the signature information gain based on the snapshot of the current round, the current set of competing candidates, and the fixed basic matching strength within the current candidate generation. Only high-query-value terms that have not yet been evaluated, or terms newly added due to the expansion of the competing candidate set, are selected. Terms that have already been evaluated and have obtained complete relations are not requested repeatedly, ensuring that the model call budget is monotonically consumed with each round without repeated allocation. When the query values ​​are equal, a deterministic total order is formed by ascending order of immutable term identifiers, enabling deterministic recalculation of the term selection results in the same state.

[0087] It should be understood that the distinguishing term is essentially a quantity defined relative to the current set of competing candidates, rather than an inherent attribute of the text itself. The distinguishing term refers to a term that can substantially change the distribution of candidates in the current set of competing candidates. The criterion is neither the rarity of the term in the corpus nor the matching strength between the term and a particular candidate, but rather whether the value distribution of the term on the signatures of each competing candidate can separate this group of candidates. Therefore, the same term acquires different query values ​​in different competitive scenarios: when competing candidates share the same hierarchical system, terms expressing the hierarchical level fall into the same signature bin for all candidates, resulting in zero signature information gain; however, when the difference between competing candidates lies precisely in the location, terms expressing the location split the competing candidates in two, resulting in a positive signature information gain.

[0088] Since the query value is calculated based on the current candidate distribution, this invention avoids two types of mismatches inherent in static weighting schemes. First, static weights pre-configure fixed values ​​for terms based on the entire database statistics, systematically underestimating a term that is high-frequency across the entire database but constitutes the only difference between the current two candidates. Second, static weights cannot change with the candidate set; the same weight table yields the same word selection results in two-candidate competition and five-candidate competition, while the latter obviously requires more distinguishing dimensions. Because the word selection quantity in this invention is calculated jointly by the snapshot, the competing candidate set, and the fixed basic matching strength, the word selection results change synchronously with the competitive relationship, allowing the limited model call budget to be allocated to terms that can change the relative order of candidates.

[0089] Let's illustrate with a specific scenario. Suppose that after recalling a target mention in an image examination report, two candidate concepts are obtained. Their hierarchical components are identical, but their location components differ. In the segmented terms, the term expressing the hierarchical component strongly matches both candidates and falls into the same signature bin, with conditional entropy equal to unconditional entropy and signature information gain of zero. The term expressing the location component strongly matches the first candidate but not the second, falling into different bins, with conditional entropy dropping to zero and signature information gain reaching its maximum value in the competing candidate set. If the normalized probability of the snapshot within the competing candidate set is... and Then the unconditional entropy is approximately Knight, the signature information gain of hierarchical words is zero, while the signature information gain of participle words is approximately Knight. When selecting in descending order of query value, location terms are included in the distinguishing terms while hierarchical terms are excluded. Therefore, this round of model calls only carries the judgment that can change the relative order of candidates. If a static weight scheme is used instead, hierarchical terms may receive higher fixed weights and be sent out first because they appear less frequently in this medical field. The resulting relationship is completely offset in the subsequent centralization, and a single model call is completely consumed without producing any ranking effect.

[0090] Example 3

[0091] This embodiment further refines the step in Embodiment 2, which involves "calculating the signature information gain of each term using the normalized probability of the snapshot within the competitive candidate set and the signature binning." This feature includes:

[0092] S3.2.1: The normalized probability of the snapshot within the competing candidate set is used as the distribution of the candidate random variable;

[0093] S3.2.2: The signature bins of each competing candidate on the same term are used as the signature random variables of that term;

[0094] S3.2.3: The signature information gain is determined as the difference between the entropy of the candidate random variable and the conditional entropy of the candidate random variable under the signature random variable condition;

[0095] S3.2.4: Calculate the conditional entropy by grouping the competing candidate set according to the signature bins, and the signature bins with a probability mass of zero are not included in the summation of the conditional entropy;

[0096] S3.2.5: When all candidates in the competing candidate set have the same signature binning on the same term, the signature information gain of that term is zero;

[0097] S3.2.6: When the numerical error causes the signature information gain to be negative, the signature information gain is clipped to zero.

[0098] Let candidate random variables be defined. Follows the distribution defined by equation (6) Signature random variable The signature information gain is given by the signature bin, as shown in equation (12):

[0099] Equation (12):

[0100] in, Represents Shannon entropy. This represents the conditional entropy of the candidate random variable under the conditions of the signature random variable. Indicates terms The signature information gain. Conditional entropy according to For deterministic binning of competing candidate groups, the contribution of a bin with a probability quality of zero is zero.

[0101] The principle behind using mutual information rather than matching strength as the basis for word selection is that terms with high matching strength do not necessarily have discriminative power. If a term strongly matches all competing candidates, it falls into the same bin, and its conditional entropy equals its unconditional entropy. If the conditional entropy is zero, the model's judgment on it, regardless of the relationship it returns, will be canceled out in subsequent centralization, making a model call for it ineffective; conversely, splitting the competing candidate terms in two makes the conditional entropy significantly lower than the unconditional entropy. Taking a positive value allows the relative order of candidates to be changed once the relationship is returned. Because... The calculation relies only on the same snapshot, the current set of competing candidates, and the fixed basic matching strength and binning threshold within the candidate generation. There is no data loop between the word selection process and the relationship feedback process, and deterministic recalculation is possible under the same state. The theoretical range of values ​​is zero to... The small negative values ​​caused by numerical errors are clipped to zero according to S3.2.6 to prevent negative gains from being mistakenly counted as subtraction terms in the combination function.

[0102] It should be noted that the entropy normalization process used in this invention is different in two places: process monitoring uses entropy normalized to the number of candidates to compare the degree of uncertainty in different rounds within the same request; while the signature information gain uses unnormalized Shannon entropy because conditional entropy and unconditional entropy must be subtracted under the same dimension, and normalization would make the gains under different competing candidate numbers incomparable. The contribution of terms with zero probability to entropy is taken as zero by convention.

[0103] In other embodiments, the selection quantity for distinguishing terms, in addition to the signature information gain, can also take the form of the variance of the signature values ​​of competing candidates on the same term, the expected change in pairwise ranking loss, or the submodal coverage target. Optionally, regardless of the form used, the selection quantity must be calculated only from the current finite candidate set and the observable candidate signatures, and must not reference relation values ​​that have not yet been returned or text generated by the model, so as to avoid data loops between the term selection process and the relation feedback process, which would undermine deterministic recalculation.

[0104] Example 4

[0105] This embodiment further refines the statement in Embodiment 1 that "after a closed-set identification request, the generative language model only returns the enumeration relation of the distinguishing terms to the competing candidate set, and converts it into a bounded signed relation value after integrity verification." This feature includes:

[0106] S3.6: Generate an order-independent summary of the candidate set based on the standard terminology database version and the candidate set ordered according to the immutable candidate identifier specification;

[0107] S3.7: Generate a sequence-related candidate request summary based on the span mentioned in the target, the distinguishing terms and the span of the distinguishing terms, the set of competing candidates arranged in the order of this presentation and the version information;

[0108] S3.8: Bind the closed set identification request and the response to the closed set identification request to the candidate set summary, the candidate request summary and the relation domain identifier simultaneously, and limit the value of the enumeration relation to three enumeration values: support, exclusion and irrelevant.

[0109] S3.9: Perform format, request binding, membership, uniqueness and privilege escalation checks on the response. If the response is missing any of the requested terms / candidate pairs, the entire response shall be deemed invalid.

[0110] S3.10: For the same term-candidate pair, use at least two equivalent requests with different candidate presentation orders. Each equivalent request is bound to a corresponding candidate request summary. If the enumeration relations returned by each equivalent request are inconsistent, the term-candidate pair is judged as unobserved, and the relation row containing the term-candidate pair is not included in the subsequent centralization.

[0111] The division of labor between the two summaries is a key structural feature of this embodiment. The input to the candidate set summary is the terminology database version and the complete candidate set sorted according to the immutable candidate identifier specification. Therefore, its value remains unchanged even if the candidate members remain the same but the presentation order changes. It identifies "which set of candidates is currently being faced," defining the candidate generation and serving as a criterion for whether the candidate set coverage probability remains read-only. The input to the candidate request summary includes competing candidates arranged in the current presentation order, distinguishing terms and their spans, and version numbers. Its value changes if any of these elements change. It identifies "which specific relational mesh request this is," used to block late responses, cross-request responses, and tampered responses. The request identifier, round number, and random number are validated separately as transaction fields and are not written into the content summary that can be reused across requests.

[0112] The reason for applying equivalent requests with different candidate presentation orders to the same term-candidate pair and classifying them as unobservable when inconsistent is that the output of generative language models is sensitive to the candidate presentation order, and this sensitivity cannot be completely eliminated through cueing engineering. Since the two equivalent requests are each bound to a candidate request summary reflecting their respective order, and their results are required to be consistent for them to be considered reliable, unstable relations caused by order sensitivity are identified as unobservable before entering the cumulative relation set. Furthermore, since any unobservable term makes the entire relation row incomplete, thus preventing the entire row from participating in centering, unstable relations will not infiltrate updates as partial rows, thereby solving the problem of model output sensitivity to candidate presentation order in the background technology. This mechanism detects and blocks observed order inconsistencies without changing the model's own order sensitivity. In other implementations, consistency detection can also be achieved by using multiple independent calls under the same order and deduplicating them idempotently by source key.

[0113] Specifically, the closed-set identification request is carried by a structured payload, whose fields include at least the schema version, request identifier, mapping unit identifier, round number, random number, input summary, candidate set summary, candidate request summary, relation domain identifier, equivalence group identifier, and output schema identifier, as well as the standard terminology base version, hint version, model version, and relation calibration version. The fact area carries the surface text and span of the target mention, the term area lists the identifier, surface text, and span of each distinguishing term, the candidate area lists the immutable identifier and brief description of each competing candidate in the order of presentation, and the relation area limits the output to a quadruple array consisting of term identifier, candidate identifier, enumerated relation, and reason code. The response must echo the above-mentioned bound fields as is, wherein the request identifier, round number, and random number are validated separately as transaction fields and are not written into the content summary that can be reused across requests.

[0114] In this embodiment of the invention, the verification is performed in eight steps in a fixed order. Syntax verification requires that the output be strictly parsable, and that the field types, enumeration values, array lengths, and required fields conform to a preset pattern. Request binding verification requires that each version number, two summaries, and relation domain identifier be consistent with the current non-expired request; equivalent requests with candidate reordering use the same equivalence group identifier and each has a candidate request summary reflecting its respective order. Term membership verification requires that each term identifier exist within the distinguishing terms in this round, and checks whether the surface text and span already bound in the request still belong to the current input summary. Candidate membership verification requires that each candidate identifier belong to the current competing candidate set; creating or rewriting candidate identifiers is prohibited. Uniqueness and completeness verification requires that the same term-candidate pair appear at most once and cover all requested term-candidate pairs. Cause compatibility verification requires that the cause code be compatible with the candidate's structured label and original scope; for example, when the cause code points to the same location, the candidate's structured location must not explicitly conflict with the input; when the cause code points to original negation, the corresponding term must be within the identified negation scope. Unauthorized content validation requires that the output does not contain the final mapping field, free terms outside the candidate set, status judgment suggestions, or replacements of the original entity. Numerical and scale validation requires that the relation value, the number of relation entries, the character length, and the processing latency are all within the preset range.

[0115] Preferably, when syntax validation fails, a single retry to fix only the format is allowed under the same candidate and the same term; this retry consumes a separately registered format repair budget. However, when membership, span, or unauthorized validation fails, the entire feedback round is invalidated, and the generative language model is not allowed to correct the candidate set itself. Missing term-candidate pairs are not considered correctable format errors and therefore do not consume format repair budget. All validations are performed by rules or code, and are not self-verified by the same generative language model to avoid inheriting the homology bias of the validated object in the validation process.

[0116] Each validated relational observation is bound to a source key consisting of the model version, prompt version, template identifier, equivalence group identifier, candidate presentation order identifier, call identifier, and the candidate request summary. Duplicate responses with the same source key are retained only once in an idempotent manner; if responses with different content appear for the same source key, the entire response from that source is invalid. For multiple valid sources of the same term-candidate pair, they are first deterministically sorted by source key, and then judged according to a pre-registered conflict threshold: if the original enumeration relations returned by the candidate reordering equivalence requests are inconsistent, they are directly judged as unobserved; if the discrepancies between other sources exceed the conflict threshold, they are also judged as unobserved; only when the threshold is not reached is a deterministic merge performed once from the original valid observations.

[0117] The dual-summary binding and integrity verification refer to a set of structural constraints that use two content summaries with different scopes to lock the candidate identity and the request identity respectively, and deprive the generative language model of its implicit selection rights by using a dense Cartesian product coverage requirement. The candidate set summary is bound to the identity rather than the expression, and its input is ordered according to the immutable candidate identifier specification, so the summary does not change with the presentation order when the candidate members remain unchanged; the candidate request summary is bound to the complete form of the current request, and any change in terms, span, candidate order or version number will change its value. Both, along with the relation domain identifier, are written into the request and a response is required, constituting the only entry point for the model output to enter the deterministic module.

[0118] The technical advantage of using a dual-summary approach instead of a single-summary approach is that the two types of failures are intercepted separately without masking each other. If only a single order-sensitive summary is used, equivalent requests that present only a change in order but not a change in the candidate set will be judged as different candidate sets, the candidate generation will be unnecessarily advanced, the candidate set coverage probability will be re-estimated, and the probability isolation mechanism will be rendered ineffective. If only a single order-independent summary is used, late responses and cross-request responses cannot be identified because the summaries are the same, and the relationship of one request may be misused for another request. Since the two summaries each bear the responsibility of defining the candidate generation and blocking request mismatches, read-only constraints and response identification can be established simultaneously. The technical effect of the integrity check is to eliminate the model's silent selection: if the model is allowed to return only a few terms—candidate pairs—that it is more confident in, the unreturned parts essentially express a judgment that is not subject to check, and the model regains the decision-making power through selective silence; the dense coverage requirement makes this silence unable to pass the check, and missing items are all invalidated.

[0119] Let's illustrate the operation of this mechanism with a specific scenario. Suppose a request submits three distinguishing terms and two competing candidates to the generative language model. The complete relation mesh should contain six terms—candidate pairs. If the response returns only five pairs, omitting one, the uniqueness and completeness checks fail, the entire round of feedback is invalid, the mapped data state remains at the pre-update checkpoint, and this omission does not consume format repair budget. Further suppose the response is complete in terms of fields, but one of the candidate identifiers does not belong to the current competing candidate set, but is obtained by the model itself. In this case, the candidate membership check will reject the response before the parameter update, as the out-of-bounds identifier has no path to the bounded signed relation value. Finally, suppose a response returns only after the next round of requests due to network latency. Its candidate request digest and random number both belong to the previous round and are inconsistent with the current round; this response is discarded directly. However, equivalent requests initiated after the same batch of candidates are swapped retain the same candidate set digest, so the candidate set coverage probability is not re-estimated due to this reordering.

[0120] Example 5

[0121] This embodiment further refines the "converting to bounded signed relation values ​​after integrity verification" in Embodiment 4. This feature includes:

[0122] S3.10.1: Use a relation validation set that is separate from the candidate ranking training set and has manually labeled terms and candidate relations;

[0123] S3.10.2: Create relation calibration groups based on one or more of the following: model version, prompt template, medical domain, relation type, and reason code;

[0124] S3.10.3: Based on the conditional confusion matrix of the enumerated relation output by the generative language model in the relation calibration group relative to the manually labeled relation, the enumerated relation is converted into a signed expectation and used as the bounded signed relation value;

[0125] S3.10.4: When the sample size of the relationship calibration group reaches the registered minimum sample size and the lower bound of the confidence interval of the registered relationship calibration performance index reaches the registered activation threshold, the relationship calibration group is activated; otherwise, it is deactivated if the pre-registered cause code, medical field and prompt template hierarchy is returned and the conditions are still not met.

[0126] S3.10.5: The transformation does not use the confidence level reported by the generative language model, and the statistical data of the conditional confusion matrix is ​​not combined across model versions without independent verification.

[0127] Let the enumeration relation output by the model be... The relationship between manual annotation is Then the bounded signed relation value is as shown in equation (13):

[0128] Equation (13):

[0129] in, Representation model for terms With candidates The returned enumeration relation, The conditional confusion matrix is ​​estimated by additive smoothing. Because it takes the value of a convex combination of three bounded values, it must be located in .

[0130] The use of an externally labeled conditional confusion matrix instead of model self-reported confidence stems from the different objects they measure: self-reported confidence is a subjective statement by the model about its own output, and its value has no verified correspondence with the empirical frequency of that output being true; while the conditional confusion matrix measures the "empirical frequency of manually labeled support when the model says it supports it," a measurable external quantity on an independent labeled set. Since the transformed relation value equals the expectation under that empirical distribution, the relation strength is converted into a comparable, summable, and numerically bounded dimension, allowing relations from different model versions and different prompt templates to participate in subsequent centering and projection on the same scale. This solves the problem in the background technique where model self-reported confidence is accepted without external calibration. The dual thresholds of sample size and lower bound of the confidence interval prevent small sample groups from giving artificially high calibration performance; hierarchical backoff gradually relaxes grouping by cause code, medical domain, and prompt template within the same model version. If backoff still does not meet the requirements, the group is disabled and the corresponding relation is not updated. The relation calibration set must be separated from the candidate ranking training set; otherwise, the ranker and relation calibrator will share the same batch of label biases.

[0131] Specifically, the conditional confusion matrix is ​​estimated using additive smoothing to avoid zero-frequency units appearing in small sample groups, which would cause the signed expectation to reach the endpoint of the interval. The relation calibration performance index can be a three-class balanced accuracy or a sign consistency rate, but only one index is registered and used in the same version to avoid artificially inflating the activation ratio by selecting the best index among different groups. The activation criteria for the relation calibration group are that the sample size reaches the registered minimum sample size and the lower bound of the confidence interval of the index at the registered confidence level reaches the registered activation threshold; both of these conditions must be met simultaneously.

[0132] Furthermore, a relation gating value is set for each relation observation: the gating value is one only when the schema, members, span, reason code, version number, and the relation calibration group are all valid; otherwise, it is zero. Relationships gated to zero do not participate in the subsequent calculation of candidate residual target values. Optionally, the dimensions for establishing the relation calibration group include model version, prompt template, medical domain, relation type, and reason code. Hierarchical fallback within the same model version gradually relaxes the grouping granularity in the order of reason code, medical domain, and prompt template. Without independent cross-version verification, statistical data from different model versions should not be merged because the output distributions of different versions are inherently incomparable.

[0133] Example 6

[0134] Based on the foregoing embodiments, the method further includes the determination and branch handling of coverage alarms. For example... Figure 3 As shown, the process includes:

[0135] Step A: Before the reconstruction of the centralization, the shared term weights, and the candidate residuals, the weighted proportion of complete relation rows in the competitive candidate set that are mutually exclusive is calculated using the initial shared term weights and the relation reliability independent of the sign of the relation value as statistical weights, to obtain the common exclusion index.

[0136] Step B: Calculate the weighted proportion of complete relation rows that exclude each candidate using the same statistical weight, and take the minimum value for all candidates in the competitive candidate set to obtain the dispersion exclusion index;

[0137] Step C: When the common repulsion index reaches the first threshold or the dispersed repulsion index reaches the second threshold, a coverage alarm is triggered;

[0138] Step D: When the coverage alarm is triggered, the bounded signed relation values ​​of this batch are written only to the audit record, and the shared term weight, the candidate residual, the candidate score, the candidate conditional probability and the candidate set coverage probability are kept to their previous values.

[0139] Step E: When the coverage alarm is triggered and there are candidates in the candidate set that have not entered the competing candidate set, expand the competing candidate set and establish a new relation domain; otherwise, perform supplementary recall when the supplementary recall budget allows; if the supplementary recall budget is insufficient, output an unmapped state or a manual review mark.

[0140] The definitions of the two indicators are shown in equations (14) and (15):

[0141] Equation (14):

[0142] Equation (15):

[0143] in, This represents the set of terms corresponding to all complete relation rows within the current relation domain. This represents the set of terms corresponding to a complete relation row where all candidates in the competing candidate set are mutually exclusive. Indicates exclusion of candidates The complete set of terms corresponding to the relation row. Indicates terms The reliability of the relationship in the row containing the relation. Indicators of common exclusion This indicates a dispersion exclusion indicator. Correspondingly, the trigger condition is... or ,in Indicates the first threshold. This indicates the second threshold.

[0144] The reason for using two metrics instead of one is that recall and omission present two forms in the relation matrix. When the correct concept is not recalled but a decisive term exists in the original text, that term rejects all competing candidates; this form is... Capture; however, when several terms exclude different candidates respectively, and no single term excludes all candidates simultaneously. The value is always zero; this form can only be obtained by accumulating the values ​​of each candidate and then taking the minimum value. capture-- Reaching the threshold means that every competing candidate has been excluded by some reliable terms, i.e., the entire group of candidates is negated one by one. Since both indicators are calculated before centralization and parameter updates, and the batch relationship only enters the audit without changing any parameters or probabilities upon triggering, the negative evidence of "the entire group of candidates is incorrect" is not transformed into a relative advantage of a candidate within the group through centralization—centralization only retains the difference components between candidates. Without this pre-judgment, the common negative component corresponding to the exclusion of all candidates would be eliminated and become invisible during centralization, thus solving the problem in the background technology of selecting the first candidate from the existing candidates when the correct concept has not been recalled. Threshold Registration must be based on the number of competing candidates and validation data: if the exclusion targets of several terms form non-overlapping groups, then... The upper bound is limited by the number of groups; setting it arbitrarily outside the weight distribution will lead to missed or false alarms.

[0145] In one example, the initial shared term weights of the two complete relation rows are respectively and The reliability of the relationship is 0. The first row excludes candidates. And support The second row excludes candidates. And support At this point, no row simultaneously rejects two candidates, therefore... ;and ,right and The cumulative rejection rates were respectively and Therefore .like If the common exclusion index is zero, a coverage alert will still be triggered. The relationship in this batch will only be entered into the audit log, and the system will then expand the competing candidate set or perform a supplementary recall.

[0146] Furthermore, the branch processing after triggering the coverage alert follows a fixed order of expansion, recall, and verification. The system first checks if there are still feasible candidates in the candidate set that haven't entered the competing candidate set. If so, the competing candidate set is expanded and a new relation domain is established, without consuming supplementary recall budget. If not, supplementary recall is performed within the allowable range of the supplementary recall budget. The supplementary recall query uses only the target mention and its binding context verified by fact boundaries, scope, and target identifiers. It does not use terms from other targets in the same segment, nor does it use text generated or rewritten by the generative language model, to prevent the recall scope from being negatively influenced by the model output. After supplementary recall, a new candidate generation is established, the relation domain and feedback round are reset, the candidate set, initial shared term weights, basic matching matrix, and score are reconstructed, a new candidate set summary is generated, and a new checkpoint is established. All relationships within the old candidate generation and old relation domain are invalidated and only retained in the audit log.

[0147] The aforementioned coverage alert refers to a pre-judgment criterion executed before parameter updates, taking the exclusion structure of the relation matrix as input and blocking submissions in the current batch as output. It measures not the merit of a particular candidate, but whether the entire group of candidates might be incorrect as a whole. This criterion must be placed before centralization because centralization systematically erases the evidence it relies on: centralization only retains the difference components between candidates, and when all competing candidates are collectively excluded by the same batch of terms, these exclusions constitute a common component with the same value for each candidate, which, after subtracting the weighted average, is strictly reduced to zero. In other words, without a pre-judgment criterion, strong negative evidence pointing to the entire group of candidates being incorrect would be zeroed out upon parameter updates, and the original score would still be used to select the first candidate and submit it, thus leaving no trace of the erroneous submission.

[0148] Instead of a single metric, we use two indicators, common exclusion and dispersed exclusion, to correspond to two different presentations of missed recalls on the relation matrix. In the first scenario, there is a decisive term in the original text that rejects all competing candidates, and the common exclusion metric immediately increases. In the second scenario, there is no such single term; several terms each reject different candidates, and no row simultaneously rejects all candidates. In this case, the common exclusion metric remains zero. Only the dispersed exclusion metric, which calculates the minimum value based on the cumulative rejection proportions of each candidate, can reflect the fact that each competing candidate has been refuted by some reliable evidence. Both indicators share the same statistical weight, which is independent of the sign of the relation value, making the indicators comparable and the threshold identifiable.

[0149] The resulting technical effect can be illustrated by comparing it with a scheme that does not set this criterion. Suppose a correct concept fails to enter the candidate set due to insufficient recall channel coverage, and two candidates in the candidate set conflict with the original text in two dimensions: location and grading. In a scheme without this criterion, the exclusion relationship returned by the model, after centering, only retains the difference component between the two candidates. The candidate with less conflict gains a relative advantage, and the conditional probability within the set becomes sharper. If this scheme uses the sharpness of the distribution within the set as sufficient evidence for recall, erroneous submissions will appear with high confidence and are difficult to identify from the results themselves afterward. In this invention, however, both indicators are calculated before centering. If either reaches its corresponding threshold, the coverage alarm is triggered. The relationship in this batch is only written to the audit record. The shared term weight, the candidate residual, the candidate score, the candidate conditional probability, and the candidate set coverage probability all remain at their previous values. The system then expands the competing candidate set or supplements the recall, thereby transforming a potential high-confidence erroneous submission into an explicit recall remedy. It should be understood that the final judgment on whether the correct concept is within the candidate set is still undertaken by the deterministic coverage calibrator. The common exclusion or dispersed exclusion formed by the model relationship can only trigger the coverage alarm, and cannot directly determine the final mapping result. This is fundamentally different in function from the scheme in which the model outputs negative conclusions outside the candidate set and makes a final judgment accordingly.

[0150] Example 7

[0151] This embodiment further refines the "using the initial shared term weight and the relation reliability independent of the sign of the relation value as statistical weights" in Embodiment Six. This feature includes:

[0152] Step A.1: For each complete relation row, determine the relation calibration reliability factor, source reliability factor, and structural verification reliability factor respectively;

[0153] Step A.2: Each reliability factor is given by a registration function bound to the model version, hint version, relation calibration version and structure verification rule version, and takes a value between zero and one, independent of the sign of the bounded signed relation value;

[0154] Step A.3: The minimum value among the relationship calibration reliability factor, the source reliability factor, and the structural verification reliability factor is taken as the relationship reliability;

[0155] Step A.4: Use the product of the initial shared term weight of the term corresponding to the complete relation row and the relation reliability as the statistical weight;

[0156] Step A.5: When the sum of the statistical weights of all complete relation rows is zero, the closed set relation feedback in this round is determined to be invalid, and the candidate condition probability and the candidate set coverage probability are kept at the values ​​before the update.

[0157] The reliability of the relationship is shown in equation (16):

[0158] Equation (16):

[0159] in, The relationship calibration reliability factor depends on whether the relationship calibration group is enabled and its registration performance. The source reliability factor depends on the registration level of the model version and the prompt version. The structure verification reliability factor depends on the compatibility between the cause code and the candidate structured label.

[0160] Using the lower bound of the three factors instead of their product or mean ensures that the failure of any one component independently reduces the statistical weight of that row without being diluted by the high scores of the other components. Even if a relation comes from the highest-ranking model version, its credibility as evidence for a coverage alert should not exceed the level allowed by that inactive state, provided its relation calibration group is not enabled. The reliability factor must be independent of the sign of the relation value; otherwise, using relation symbols in reliability construction will give exclusionary relations different evidentiary weights than supporting relations. This would cause the triggering of coverage alerts to depend on the relation direction itself rather than the quality of evidence, thus rendering the threshold identifiable.

[0161] The reliability factors are provided by the registration function bound to the corresponding version: the relation calibration reliability factor depends on whether the relation calibration group is enabled, hierarchical rollback, or disabled, and its registration performance level; the source reliability factor depends on the registration level of the model version and the prompt version; the structure verification reliability factor depends on the compatibility between the reason code and the candidate structured label. Optionally, if any related version is not registered, the corresponding factor is zero, and the relation row will not be weighted in the statistics of the coverage alarm and will be simultaneously excluded from the subsequent calculation of the candidate residual target value.

[0162] Example 8

[0163] This embodiment further refines the step in Embodiment 1, which involves "centering the complete relation according to the normalized probability of the snapshot within the competitive candidate set to obtain candidate relative feedback with a weighted sum of zero." This feature includes:

[0164] S4.1: For the bounded signed relation values ​​accumulated up to this round within the current relation domain, determine the cumulative observation mask according to the term-candidate pair;

[0165] S4.2: When all candidates in the competing candidate set have been observed for the same term, the relation row corresponding to the term is determined as a complete relation row. Relation rows with unobserved term-candidate pairs are set to zero and do not participate in the centralization.

[0166] S4.3: For each complete relation row, calculate the weighted average of the bounded signed relation values ​​of the complete relation row using the normalized probability of the snapshot in the competing candidate set, and subtract the weighted average from the bounded signed relation values ​​of each candidate in the complete relation row to obtain the candidate relative feedback;

[0167] S4.4: When the bounded signed relation values ​​of the same complete relation row for each candidate in the competing candidate set are the same, the relative feedback of the candidate in the complete relation row is zero;

[0168] S4.5: In each feedback round, the snapshot of this round is used to uniformly re-centralize all complete relations accumulated in the current relation domain, without centralizing the already centralized results again;

[0169] S4.6: When the members of the competing candidate set change, a new relation domain is established, the feedback rounds in the new relation domain are reset to zero, and the bounded signed relation values ​​of the previous relation domain do not participate in the centralization in the new relation domain.

[0170] Cumulative observation mask A value of 1 indicates that the corresponding term—candidate pair—is still observed after merging the original observations and full departmental control; otherwise, a value of 0 is used. The criterion for a complete relation row is... .

[0171] The reason why the uncentralized original relation values ​​are preserved in a multi-round feedback system and uniformly re-centralized in each round using a snapshot of that round, instead of accumulating the centralization results of each round, is because the benchmark for centralization... The nature of the feedback changes with each round: if the relative feedback in the first round is calculated using the first round's benchmark, and the relative feedback in the second round is calculated using the second round's benchmark, the sum of the two is no longer weighted zero with respect to either benchmark, and the cancellation property of the common components immediately fails. Since this embodiment always preserves the original observations and re-centers the data from the original values ​​in each round, regardless of the number of feedback rounds, the quantities involved in the update are always weighted zero with respect to the benchmark of the current round, and the cancellation property remains valid across multiple rounds. The isolation of the relational domain is similar: once the members of the competing candidate set change, The defined probability space has changed, and the values ​​of the old relation rows for the new competing candidates have not been observed. If cross-domain mixing is used, the integrity determination itself loses its meaning. Therefore, the new relation domain must re-request the complete Cartesian product of the retained terms and the new competing candidates. The old relation is only used for auditing.

[0172] Specifically, the value of the cumulative observation mask is determined by the deterministic merging result of the original observations and the full gate control: if a term-candidate pair has at least one observation that passes all verifications and gate control, and is not judged as unobserved due to inconsistency of equivalent requests or source conflicts exceeding the threshold, the corresponding mask is set to one; otherwise, it is set to zero.

[0173] The centralization described above, which uses complete relation rows as units and pre-update probability snapshots as weights, refers to a deterministic transformation that converts relation values ​​into relative feedback. The input to this transformation is the absolute relation value returned by the model and calibrated; the output is the deviation of each candidate within the same row from the weighted average of that row. Its effect is equivalent to stripping away the component that affects all competing candidates in each row, retaining only the component that changes the relative order between candidates. The fixed-point property of this transformation is that for any relation row where all competing candidates have the same value, its output is always a zero vector.

[0174] This property is precisely the mechanism by which this invention eliminates the bias of shared terms. In medical texts, location terms, grade terms, and degree terms are largely shared by multiple standard concepts. Static weighting schemes and schemes that directly superimpose model scores both allow the supporting components of these terms to simultaneously enter the scores of each competing candidate. The consequences are not merely ineffective computation: because the base scores of each candidate are different, the same common component is non-linearly amplified or compressed after subsequent probability calibration. Shared terms that were originally indistinguishable actually change the probability interval between candidates, contaminating submission criteria such as the first-to-last interval. Because this invention centers the terms before any parameter update, the contribution of shared terms is strictly reduced to a zero vector rather than approximately zero, thus structurally eliminating the aforementioned contamination of the submission criteria.

[0175] Using complete relation rows, rather than individual term-candidate pairs, as the centering unit addresses another type of failure. If relation rows with unobserved terms are allowed to participate in centering with default values, the unobserved terms will be implicitly included in the weighted average as irrelevant relations, shifting the centering benchmark of that row. The relative feedback of the remaining observed terms' candidate pairs will be systematically pulled to one side. Since unobserved terms are often not random but concentrated on term-candidate pairs where the model's judgment is unstable, this shift is related to the judgment difficulty and cannot be eliminated by increasing the number of samples. This invention sets all relation rows with unobserved terms to zero, preferring to discard some of the existing information in that row rather than allowing a biased benchmark to enter the update.

[0176] Consistency under multi-round feedback is guaranteed by re-centralization rather than round-by-round accumulation. Since the centralization benchmark itself changes with each round, if the candidate relative feedback in the first round is calculated using a snapshot of the first round, and the candidate relative feedback in the second round is calculated using a snapshot of the second round, the sum of the two is no longer weighted zero with respect to any snapshot. The cancellation property of common components then fails, and the accumulated common components will re-form a bias after multiple rounds. This invention always preserves the original, uncentralized relational observations. In each round, a unified re-centralization is performed on all accumulated complete relations within the current relational domain using the snapshot of that round. Therefore, the quantities participating in the update are always weighted zero with respect to the benchmark of that round, and the cancellation property remains valid regardless of the number of rounds.

[0177] For example, suppose there are two complete relation rows accumulated in a certain relation domain, one of which corresponds to a hierarchical word shared by two competing candidates. The model returns support for both candidates, and the calibrated bounded signed relation values ​​are the same. The other row corresponds to the part words that only match the first candidate part, and the calibrated bounded signed relation values ​​are respectively and The normalized probabilities within the competitive candidate set. and After centralization, the weighted average of the relation rows containing the hierarchical terms is... The relative feedback between the two candidates is zero, and this row does not provide a relative advantage to any candidate; the weighted average of the relation rows containing the part words is approximately The relative feedbacks of the two candidates are as follows: and The weighted sum is zero. If new relation rows are added in the second round because the competing candidates remain unchanged, the system does not add the results of the first round to the results of the second round. Instead, it uses the snapshot of the second round to uniformly re-center all complete relation rows accumulated in the two rounds. The relation rows containing the hierarchical words still output zero vectors under the new benchmark.

[0178] Example 9

[0179] This embodiment further refines the step in Embodiment 1, which involves "reconstructing the shared term weights and sign-preserving candidate residuals from initial values ​​within the item-by-item boundaries and the total budget." For example... Figure 4 As shown, this feature includes:

[0180] S4.7: Determine the distinguishing strength of each term by the weighted L2 norm of the normalized probability of the snapshot within the competing candidate set based on the relative feedback of the candidate terms;

[0181] S4.8: The initial shared term weights are multiplicatively updated and normalized according to the distinguishing strength, and then projected onto the feasible region that simultaneously satisfies the weight sum of one, the multiplier box constraint relative to the initial shared term weights item by item, and the first norm total budget constraint, to obtain the reconstructed shared term weights.

[0182] S4.9: The product of the candidate relative feedback, the non-negative update step size, and the relation validity gate value is used as the candidate residual target value;

[0183] S4.10: Project the candidate residual target value onto the feasible region that simultaneously satisfies the terms-by-term absolute value boundary, the first norm total budget, the residual terms not belonging to the complete relation row being zero, the residuals within the complete relation row having zero mean weighted by the normalized probability, and the sign-preserving constraint, to obtain the reconstructed candidate residual;

[0184] S4.11: The sign-preserving constraint is that the product of the reconstructed candidate residual and the corresponding candidate relative feedback is non-negative, and the candidate residual is zero when the candidate relative feedback is zero;

[0185] S4.12: If any of the assertions in the following categories after projection—the absolute value boundary, the norm total budget, the weighted zero mean, and the sign preservation constraint—fail, the current update will not be committed and the process will be rolled back to the checkpoint before the update.

[0186] The two feasible regions are shown in equations (17) and (18), respectively:

[0187] Equation (17):

[0188] Equation (18):

[0189] in, and Let these represent the lower bound factor and upper bound factor of the multiplier box, respectively. , The total budget of the first norm representing the weights of shared terms. Denotes the term-by-term absolute value boundary of the candidate residuals. This represents the total budget of the L1 norm of the candidate residuals; the other symbols have the same meaning as before. The initial shared term weights belong to... Zero residuals belong to Therefore, both feasible regions are non-empty closed convex sets, and the projection result is unique.

[0190] The technical significance of dual-channel separation lies in the fact that shared term weights answer "how important is this term?", their values ​​are non-negative and sum to one, and cannot express direction; candidate residuals answer "whether this term supports or rejects a specific candidate", their values ​​are signed and separated from the base matching strength located in the zero-to-one interval—precisely because of this separation, when the base matching strength is already at the endpoint of the interval, the feedback of support or rejection can still be expressed through the residuals without being clipped by the interval. Weighted zero-mean constraints and sign-preserving constraints prevent two types of degradation: asymmetric pruning will reintroduce the already zero-sum relative feedback into the common component, causing the canceled shared term bias to be reproduced through projection; while unsigned projection may map positive relative feedback indicating support to negative residuals, so that even if the monotonicity of the score holds, it cannot guarantee the feedback direction. Since both are written into the same feasible region and asserted one by one after projection, any violation triggers a global rollback, and the upper bound of the feedback direction and amplitude holds structurally simultaneously. Regarding parameters, and If the value is too small, the candidate residuals will be significantly pruned in the projection, and multiple rounds of feedback will hardly change the first candidate. If the value is too large, a single relation can dominate the reordering, weakening the significance of boundedness. Since the shared term weights are non-negative and sum to one, the absolute value of the candidate directional feedback feature does not exceed the sum of the boundary values ​​of the weighted residuals of the shared term weights and the maximum shared term weight multiplied by... "The smaller of the two, therefore" , and Joint calibration should be performed to match the upper bound of the score change in a single round with the direct submission threshold.

[0191] In this embodiment of the invention, the projection can be implemented by an alternating projection, dual ascent, or a combination of solvers for box constraints and L1 norm constraints. The maximum number of iterations and convergence tolerance of the solver are pre-registered. If the solution fails to converge within the registered number of iterations and tolerance, or if the solution result still violates the term-by-term absolute value boundary, total budget, weighted zero mean, or sign-preserving constraints after independent assertion, the temporary mapped data state is discarded entirely and rolled back to the pre-update checkpoint. Approximate solutions are not allowed to be accepted by relaxing the constraints. It should be noted that the above assertions are executed by code independent of the solver to prevent the tolerance within the solver from being used as a criterion by the same implementation.

[0192] It is worth noting that the shared term weights and the candidate residuals can be stored as a product in a subordinate storage format. However, in this case, the strictly positive shared term weights must also be stored, and the candidate residuals must be restored without loss by dividing each term. The product matrix alone is insufficient to claim the uniqueness of the decomposition. Alternatively, other lossless structures that explicitly encode both magnitude and direction can be used, as long as the two restored components still satisfy the constraints of non-negativity normalization and sign preservation, respectively.

[0193] The dual-channel bounded anchoring reconstruction refers to decomposing the influence of closed-set relation feedback on the scoring input into two separable channels: amplitude and direction. The two channels are reconstructed from the same initial anchor point in each round, rather than being cumulatively multiplied from the previous round's results. The amplitude channel is the shared term weight, which is non-negative and sums to one, indicating the importance of the term in the current mapping. The direction channel is the candidate residual, which is signed and separate from the basic matching strength in the zero-to-one interval, indicating whether the term supports or rejects a specific candidate. The two channels are coupled through the candidate directional feedback feature before entering the fixed scoring function.

[0194] The necessity of channel separation can be seen from a counterexample. If independent candidate residual channels are not set up, and the signed relational feedback is directly superimposed on the basic matching strength, then when the basic matching strength between a term and a candidate has reached the upper end of the interval, further support feedback will be completely lost due to interval pruning, while the rejection feedback of the same term can still take effect, thus making the directionality of the feedback asymmetrical. Conversely, if shared term weight channels are not set up and the candidate residuals simultaneously bear the amplitude and direction, then a strong feedback can simultaneously change the importance of the term and the relative merit of the candidate, and the influence of the two cannot be separately bounded, thus weakening the meaning of boundedness. Because this invention constrains the amplitude by normalization and multiplier boxes, and the direction by term-by-term boundaries and sign preservation constraints, both types of influences have calculable upper bounds.

[0195] Reconstructing from the initial anchor point rather than incrementally accumulating corresponds to the reproducibility requirement. In the incremental scheme, the first... The parameters of the wheel depend on the first... Turn number All intermediate states of a round, and the numerical details of any round will propagate along the rounds. The same input may not yield the same result in different execution environments, and backtracking to a certain historical state also requires saving all intermediate values. In this invention, each round starts with the initial shared term weights and zero-value candidate residuals, and reconstructs parameters using the accumulated original relation observations in the current relation domain as input. Therefore, the first round... The parameters of a wheel are deterministic functions of the original set of observations and are independent of the path leading to that wheel; thus, the parameters of any wheel under the same input and version can be deterministically recalculated offline, and rollback only requires restoring the checkpoint without replaying.

[0196] The sign-preservation constraint and the weighted zero-mean constraint respectively prevent two types of degradation, and their technical effects can be compared with the default scheme. Without the sign-preservation constraint, the projection asymmetrically prunes the components to meet the total budget, potentially mapping positive candidate relative feedback representing support to negative candidate residuals. In this case, even if the fixed scoring function is strictly monotonic with respect to the candidate directional feedback features, a valid support feedback will still reduce the score of the corresponding candidate, and the feedback direction is reversed in the final step. Without the weighted zero-mean constraint, the asymmetric pruning reintroduces the already zero-sum candidate relative feedback into the common components, causing the shared term bias that was canceled out in the centering to reappear through the projection. Since the two constraints are written into the same feasible region and asserted one by one after projection, any violation triggers a global rollback. The preservation of the feedback direction and the holding of the upper bound of the magnitude become structural properties rather than training results.

[0197] Let's illustrate the role of boundedness with a numerical scenario. Let the term-by-term absolute value boundary be... The total budget for the first norm is The candidate relative feedback for a given complete relation row is and The product of the non-negative update step size and the relation validity gating value is The candidate residual objective value is and Neither of the two terms touches the absolute value boundary, and their first norm is... Also within the total budget, the projected value remains unchanged; if the relation validity gate value for the same relation row is one and the step size is... The target value then becomes and Both items cross the boundary of absolute value for each item. Projection shrinks them to within the boundary while maintaining the weighted zero mean and consistent sign. Therefore, the score change caused by a single round of feedback has an upper bound and will not be dominated by a single relation for rearrangement.

[0198] Example 10

[0199] This embodiment provides a less computationally intensive implementation of Embodiment Nine's method of "projecting the candidate residual target value onto a feasible region that simultaneously satisfies the following constraints: term-by-term absolute value boundary, norm 1 total budget, zero residual terms not belonging to the complete relation row, zero mean of residuals within the complete relation row with respect to the normalized probability weighted average, and sign-preserving constraints, to obtain the reconstructed candidate residual," including:

[0200] S4.10.1: Represent the reconstructed candidate residuals as the product of the coefficients of the same nonnegative row and the candidate relative feedback of the complete relation row;

[0201] S4.10.2: The upper limit of the non-negative row coefficient is taken as the minimum value of the ratio of the absolute value boundary of each non-zero candidate in the complete relation row to the absolute value of the candidate relative feedback;

[0202] S4.10.3: Set the coefficients of the non-negative rows that do not belong to the complete relation rows to zero;

[0203] S4.10.4: Ensure that the sum of the products of the nonnegative row coefficients of each complete relation row and the first norm of the candidate relative feedback of that complete relation row does not exceed the total budget of the first norm;

[0204] S4.10.5: The weighted sum of the candidate relative feedback with respect to the normalized probability is zero, so that the candidate residuals obtained by scaling with the non-negative row coefficients satisfy the weighted zero mean and the sign retention constraint.

[0205] The implementation method is shown in equation (19):

[0206] Equation (19): ,in ,and

[0207] in, Represents a complete relational row Non-negative row coefficients, all-zero rows take .

[0208] The reason why row-wise scaling avoids explicit projection of the weighted zero-mean and sign-preserving constraints is that these two constraints are invariant under in-row scaling: the candidate relative feedback is already satisfied. Multiplying by the same nonnegative scalar and then weighting the result still yields zero; similarly, a nonnegative scalar does not change the sign of each component, therefore... This is automatically achieved. Therefore, the projection, which originally required solving on a high-dimensional closed convex set, degenerates into a one-dimensional truncation of the row coefficient vector and total budget allocation. The computational complexity is reduced from being related to the product of the number of candidates and terms to being related to the number of complete relation rows, making it suitable for deployment scenarios with many competing candidates or tight latency budgets. The cost is that the feasible region is shrunk to a subset of the original feasible region—the residuals of each candidate within a row are forced to be distributed proportionally according to relative feedback, making it impossible to independently adjust a candidate within the item-by-item boundary. Therefore, when the item-by-item boundary frequently becomes a tight constraint, its effect is weaker than the complete projection in Example 9.

[0209] Furthermore, the allocation of the non-negative row coefficients among the complete relation rows can be proportional to the relation reliability or the first norm of the candidate relative feedback of each row, or it can be allocated sequentially in descending order of the distinguishing strength until the total first norm budget is exhausted. Optionally, this implementation method is preferred in deployment scenarios with a large number of competing candidates, a large number of terms, or a tight online latency budget; while in scenarios where the absolute value boundaries of each term frequently become tight constraints, the implementation method of complete projection is preferred to avoid the usable range of some candidates being limited by the boundaries of other candidates in the same row due to the proportional distribution within the row.

[0210] Example 11

[0211] This embodiment further refines the "obtaining candidate scores from a fixed scoring function and calibrating them as candidate conditional probabilities, and calculating candidate set coverage probabilities using only the recall and coverage features before feedback of closed-set relations" in Embodiment 1. These features include:

[0212] S2.1: The recall coverage features include at least one of the following: maximum recall score, candidate coverage rate, number of healthy recall channels, proportion of unregistered medical terms, and span quality and type conflict of the target mention;

[0213] S2.2: The recall coverage features do not include the enumeration relationship, the reconstructed shared term weights, the candidate residuals, and the candidate targeted feedback features;

[0214] S2.3: Subtract the candidate set coverage probability from one to obtain the candidate set outside probability, and use the product of the candidate set coverage probability and the candidate conditional probability to obtain the joint probability of each candidate;

[0215] S2.4: The calibration of the candidate conditional probability and the calculation of the candidate set coverage probability adopt mutually isolated data paths;

[0216] S2.5: When the candidate set summary remains unchanged, the candidate set coverage probability and the candidate set outside probability are kept read-only;

[0217] S2.6: When the candidate set summary changes due to the execution of supplementary recall, a new candidate generation is established, the candidate set summary is regenerated, and the candidate set coverage probability is recalculated with the new recall coverage features.

[0218] Since the weights of shared terms, after projection by equation (9), satisfy... Furthermore, the basic matching strength is between zero and one, and the change in text coverage has a definite upper bound, as shown in equation (20):

[0219] Equation (20):

[0220] in, Represents the candidates calculated using the initial shared term weights. The text coverage rate. This upper bound makes the impact of the feedback path on the in-set scoring input predictable, while the coverage probability path, because it does not accept any feedback as input, its value does not change with feedback within the same candidate generation.

[0221] The necessity of read-only constraints lies in distinguishing between two sources of uncertainty. In-set conditional probability answers "which one is the correct concept if it's in the set?", and its uncertainty can be reduced by relational evidence. Candidate set coverage probability answers "whether the correct concept is in the set?", and the answer to this question can only be determined by evidence from the recall phase; closed-set relational feedback, no matter how sufficient, does not carry information outside the set. Since the summary remains unchanged, i.e., the coverage probability is frozen, it is only allowed to change when supplementary recall alters the candidate set summary, enters a new candidate generation, and is re-estimated with new recall and coverage features. This ensures that "more certainty within the set" and "more sufficient recall" cannot be confused in the data path. Supplementary recall queries are constructed using only target mentions and their validated binding contexts, without using model-generated text, to prevent the recall scope from being negatively influenced by the model output.

[0222] Specifically, the coverage calibrator takes the recall coverage features as input and whether the correct concept is within the current candidate set as a binary label. It is trained with log loss on an independent calibration set containing manually constructed candidate missed negative samples and validated on a locked test set using the expected calibration error or Brier score. The candidate conditional probability calibration, on the other hand, is trained only on samples whose manually constructed truth values ​​are within the current candidate set; candidate missed samples are not included in the training set for this calibration. Therefore, the training sets for both are isolated in terms of sample composition, not just in terms of not sharing input during inference.

[0223] In this embodiment of the invention, the input to the shared monotonic calibration may include candidate score vectors and candidate set-level state features such as round number, effective relationship ratio, number of candidates, and budget usage. It is trained and locked on a calibration set that simultaneously includes initial state and feedback state replay samples. The parameters must not be changed during online feedback, nor can the initial state and feedback state be trained separately and then switched online. It should be noted that this transformation must maintain equivalence to the candidate arrangement and must not employ a candidate-by-candidate independent mapping that might reverse the score order, nor should it simply normalize multiple independent calibration outputs after nonnegation.

[0224] The two-level probability isolation refers to a probabilistic modeling approach that assigns two non-overlapping questions—whether the correct concept lies within the candidate set, and if so, which candidate it is—to two calibration pathways with disjoint input sets for separate estimation, and then uses their product to construct the joint candidate probability. The first level is answered by the candidate set coverage probability, and its input is limited to the deterministic coverage features formed during the recall phase; the second level is answered by the candidate conditional probability, and its input is the candidate scores within the set, which can accept feedback from closed-set relations.

[0225] The reason for isolation is that the feedback from closed-set relations does not contain the information needed to answer the first question. The closed-set identification request presents only the current competing candidate set to the generative language model. The model is prohibited from creating candidate labels, and the returned relations only describe the support or exclusion between a given term and a given candidate; no matter how consistent or strong these relations are, they cannot provide evidence that there are more suitable concepts outside the candidate set. If the feedback quantity or its derivatives are allowed to enter the coverage probability input, the coverage probability will increase synchronously as the in-set distribution becomes sharp due to the feedback. The scheme will arrive at the conclusion that the more repeatedly it confirms among the incorrect candidates, the more confident that the correct answer is among these candidates. The failure mode of recall omission is masked by ranking confidence.

[0226] The read-only constraint further strengthens the isolation from input disjointness to value invariance. Simply stipulating input disjointness is insufficient: some components of the coverage feature are formed by term weights and the basic matching matrix. If the shared term weights are reconstructed in the feedback and then reused in the calculation of the coverage feature, the feedback quantity will indirectly enter the coverage probability through this component. This invention stipulates that the coverage quantity allowed for the coverage feature can only be formed by the immutable initial shared term weights and the fixed basic matching matrix within the candidate generation. When the candidate set summary remains unchanged, the candidate set coverage probability and the candidate set out-of-set probability are directly frozen. Only when supplementary recall changes the candidate set summary or establishes a new candidate generation are the coverage features re-estimated with new recall coverage features, thus closing the indirect pathway.

[0227] The resulting technical effect is that the two types of uncertainty can be measured and handled separately. During the feedback process, the change in the candidate conditional probability reflects a decrease in in-set ranking uncertainty, while the unchanged candidate out-of-set probability indicates that no new evidence of recall sufficiency has been obtained. The final submission condition simultaneously checks the joint probability of the first candidate, the interval between the joint probability of the first and second candidates, and the candidate out-of-set probability; if any one of these criteria is not met, submission is not performed. Continuing with the aforementioned numerical scenario, the closed-set relation feedback causes the in-set conditional probability to change from... and Become and The joint probability is correspondingly determined by and Become and The probability of being outside the candidate set remains constant. The relative merits within the set are redistributed, but the answer to whether correct concepts are recalled remains completely untouched. If the out-of-set threshold is registered as... Even if the conditional probability of the first candidate continues to rise, the mapping still does not meet the final submission condition. The system will switch to supplementary recall instead of submission, thus preventing erroneous submissions when the correct concept has not been recalled at the probability level.

[0228] Example 12

[0229] Based on the foregoing embodiments, the method further includes state checkpointing, atomic commits, and resource ledger processing, including:

[0230] Step F: At the beginning of each feedback round, save the current accepted mapping data status as a pre-update checkpoint. The mapping data status includes candidate generation, relation domain, candidate set, competing candidate set, initial shared term weight, shared term weight, candidate residual, candidate directional feedback feature, candidate score, candidate conditional probability, candidate set coverage probability, and the cumulative uncentralized relation and data status version in the current relation domain.

[0231] Step G: When the request is consistent with the data state version and the relation integrity, numerical finiteness, probability isolation, item-by-item boundary, total budget, sign preservation and weighted zero mean assertions are all passed, the updated mapped data state is atomically committed; otherwise, the entire process is rolled back to the checkpoint before the update.

[0232] Step H: The number of feedback rounds, the number of replenishment recalls, and the number of format repairs are maintained by a monotonic resource ledger independent of the state of the mapping data. These are atomically deducted before each call to the generative language model, and the deducted resources are not restored when the state of the mapping data is rolled back.

[0233] Step I: The standard concept identifier of the first candidate is written into the medical data record to meet the final submission conditions. The final submission conditions include the joint probability of the first candidate reaching the submission threshold, the joint probability interval between the first candidate and the second candidate reaching the interval threshold, the out-of-set probability of the candidate not exceeding the out-of-set threshold, the effective relationship ratio reaching the relationship threshold, and the first candidate belonging to the completely identified set of competing candidates.

[0234] Step J: When the processing budget is exhausted, the deadline is reached, or an irrecoverable verification failure occurs and the final submission condition is not met, output an unmapped status or a manual review mark.

[0235] The effective relation ratio is defined as the ratio of the number of candidate pairs that are still observed after merging the original observations and all gates in the current main request to the total number of the deduplicated logical terms. The equivalent requests for candidate reordering are only used for gates and are not counted repeatedly, so this ratio is between zero and one.

[0236] Separating the rollbackable mapping data state from the non-rollbackable resource ledger is a key trade-off in this embodiment. If the resource count is rolled back along with the state, any failed round will refund the actual model call quota already incurred, allowing for infinite retries, thus negating the guarantee of finite iterations. Conversely, if the resource count is written outside the rollbackable state but not deducted before the call, timeout or crash calls will not be counted. Since the ledger is atomically deducted before each call using a compare-and-swap method, and rollback does not restore, each loop edge inevitably consumes a finite budget. When the budget is exhausted, the solution will inevitably enter the committed, unmapped, or reviewed final state. Atomic commit ensures that there are no intermediate values ​​in the externally visible state: shared term weights, candidate residuals, two-level probabilities, and cumulative relationships either advance to a new accepted state as a whole or remain intact at the pre-update checkpoint, thus solving the problem of erroneous content being partially written and difficult to roll back in the background technology. In other implementations, the final submission conditions may also include a group deletion sensitivity test of a fixed set of evidence: under the conditions of freezing the set of competing candidates, the snapshot before the update, the basic matching matrix and the set of candidate request summaries, delete an independent feedback group one by one and replay it from the initial anchor point. If any deletion causes a change in the first candidate or the conditional probability of the first candidate to shift beyond a threshold, the final submission is blocked.

[0237] Specifically, the thresholds for intermediate acceptance and final submission are different. Intermediate acceptance only indicates that the state obtained in this round can serve as the starting point for the next round. Its requirements include all validations, sign preservation and numerical boundary assertions passing, the candidate set coverage probability remaining unchanged, no coverage alarm being triggered, and the new first candidate having minimum independent support when the first candidate changes. Final submission, on the other hand, must also meet the final submission conditions and the first candidate stability rule. Optionally, the first round of feedback is only allowed to be submitted directly if it already contains at least two independent feedback groups, the first candidate has not changed, and all final submission conditions have been met. If the first candidate changes, it is only accepted as a state awaiting stability. Subsequent rounds must introduce new independent groups from different sources; synonymous expressions, candidate reordering, and relationships derived from the same template do not constitute independent groups.

[0238] The temporary mapping data status shall not be upgraded to the accepted status under any of the following circumstances: Structured output, identifier, span, or member validation fails; contradictory relationships are not set to unobserved according to the aforementioned rules, or the proportion of valid relationships after being set to unobserved is lower than the registration threshold; the coverage alarm is triggered, in which case only the expansion of the competing candidate set, supplementary recall, or manual review is permitted; updates exceed item-by-item boundaries, total budget of shared term weights, absolute value boundaries of candidate residuals, total budget of candidate residual norm, sign preservation constraint, or weighted zero mean constraint; the candidate set coverage probability or the... The following are considered as grounds for error: out-of-set probability of candidates being modified by feedback path, or failure to establish new candidate generations and recalibrate after supplementary recall; change of the first candidate and lack of at least two independent support groups for the new first candidate; fewer than two independent groups in the fixed evidence set at the final submission stage, change of the first candidate after deleting any group, or sensitivity exceeding the registered final threshold; model call timeout, unregistered model or prompt version, or inconsistency between the input summary and either of the two summaries; inconsistent versions of the shared monotonic calibration or inability to maintain score order, and inconsistency of non-finite values ​​or deterministic recalculation results during projection, scoring, or calibration.

[0239] Furthermore, the sensitivity test for group deletion freezes the actual evidence set involved in this reconstruction before forming a temporary update. This set is bound to the same relation domain, the same candidate request summary set, and the same pre-update snapshot. After freezing, terms, competing candidates, or evidence cannot be reselected based on any group deletion result. Evidence is divided into independent groups according to its source. Highly relevant relationships such as the same model call, the same text clue, candidate transposition equivalent requests, or homologous template derivations cannot be split into different independent groups. The re-implementation after group deletion maintains the competing candidate set, the snapshot, the basic matching matrix, the fixed candidate features, and the fixed scoring function unchanged. It re-executes merging, centering, coverage alarm determination, projection, scoring, and shared monotonic calibration from the candidate generation parameter anchor point and the zero-value candidate residual. It neither generates new evidence nor reselects competing candidates or distinguishes terms. Sensitivity is taken as the maximum offset of the conditional probability of the first candidate obtained from the recalculation of each deleted group relative to the full evidence result. If there are fewer than two independent groups, it is directly judged as unstable.

[0240] In one example, the frozen set of evidence to be submitted in the final submission contains two independent groups, and the in-set conditional probability after full evidence reconstruction is: and After deleting the first group, it becomes: and After deleting the second group, it becomes and The first candidate remained unchanged after both deletions, with a maximum offset of [value missing]. If the sensitivity threshold for registration is If the first candidate changes, the test passes; if any deletion causes the first candidate to change, the first candidate change indicator will directly prevent the final submission, even if the probability shift is small.

[0241] It is worth noting that the maximum number of feedback rounds, the supplementary recall budget, and the format repair budget are all non-negative integers at the full request level. In one implementation, the maximum number of feedback rounds is three, and the supplementary recall budget is one. Each loop edge deducts its corresponding budget before initiating a model call. Changes in candidate iterations do not restore the deducted budget; therefore, the state machine will inevitably enter the committed, unmapped, or manually reviewed final states within a finite number of steps. It should be understood that this invention requires a deterministic finite termination and does not advocate that the iterative process depends on the continuous differentiability of the objective function or necessarily reaches global convergence in a mathematical sense.

[0242] The aforementioned state checkpoints, atomic commits, and monotonic resource ledgers refer to a transaction mechanism that divides the runtime state within a request into two categories based on rollbackability and manages them separately: the mapping data state records all quantities that can be replayed and reconstructed, including candidate generations, relation domains, parameters of the two channels, two-level probabilities, and cumulative original relation observations. This type of state is rolled back as a whole when acceptance fails. The resource ledger records all consumptions that have actually occurred and are irreversible, including feedback rounds, supplementary recalls, and format repairs. This state only progresses monotonically and never rolls back.

[0243] This division represents a trade-off between finite termination and state consistency. If the resource count is rolled back along with the mapped data state, any failed round will refund the actual model call quota already incurred. Failure and retries form a closed loop, rendering the maximum feedback round limit meaningless, and requests may loop infinitely under abnormal input. Conversely, if the resource count is excluded from rollbackable states but only counted after the call returns, calls corresponding to timeouts, connection interruptions, or process crashes will not be counted, and finiteness is also violated. This invention requires that the resource ledger be atomically deducted and a ledger version generated in a comparison and exchange manner before each external call, and a call is only allowed after successful deduction. Therefore, each loop edge inevitably consumes a finite budget.

[0244] The technical effect of atomic submissions is reflected in the externally visible state. The shared term weights, candidate residuals, two-level probabilities, and cumulative relationship observations either progress to a new accepted state as the data state version advances, or remain completely at the pre-update checkpoint, with intermediate values ​​never being externally observable at any time. The audit log simultaneously saves the pre-update state version, the temporary state, and the ledger version, allowing the cause of any failure to be traced offline. In a specific scenario, if the response is missing a term-candidate pair and fails the completeness check, the entire feedback round is invalid, the mapped data state remains at the pre-update checkpoint, this omission does not consume the format repair budget, and the model calls that have already occurred are still counted in the ledger; if the projection solution does not converge within the registered number of iterations, the temporary state is discarded and rolled back, and the ledger is also not restored. Thus, the rollback of the mapping result and the limited termination of the request are simultaneously achieved.

[0245] Example 13

[0246] This embodiment provides a medical terminology standard concept mapping system, such as Figure 5 As shown, the system includes:

[0247] The mapping unit construction module is used to construct mapping units based on target references in medical texts and to recall a candidate set from a standard terminology database.

[0248] The initial scoring and two-level probability module is used to obtain candidate scores and calibrate them into candidate conditional probabilities by using the term-weighted sum of shared term weights and candidate residuals that are initially zero as candidate-directed feedback features. The candidate scores are obtained by a fixed scoring function and calibrated into candidate conditional probabilities. The candidate set coverage probability is calculated only using the recall coverage features before feedback of closed set relations.

[0249] The closed set identification and verification module is used to freeze the candidate conditional probabilities before the update as a snapshot, thereby determining the competing candidate set and distinguishing terms. After the closed set identification request, the generative language model returns only the enumeration relationship of the distinguishing terms to the competing candidate set, and after the integrity verification, it is converted into a bounded signed relation value.

[0250] The centralization and bounded reconstruction module is used to centralize the complete relation according to the normalized probability of the snapshot in the competing candidate set, to obtain the candidate relative feedback with a weighted sum of zero, so as to reconstruct the shared term weights and sign-preserving candidate residuals from the initial values ​​within the item-by-item boundary and the total budget;

[0251] The monotonic rearrangement and output module is used to re-evaluate the candidate directional feedback features by the fixed scoring function that is monotonically undiminished, update the candidate conditional probabilities through shared monotonic calibration, and output the standard concept identifier of the first candidate to be written into the medical data record.

[0252] Each of the above modules corresponds to the processing in the aforementioned method embodiments, with each detailed processing handled by its respective module. The initial scoring and two-level probability module is responsible for constructing recall coverage features, calculating the candidate set coverage probability and the candidate set out-of-set probability, and maintaining the read-only condition. The closed-set identification and verification module is responsible for calculating signature binning and signature information gain, combining query values ​​and selecting distinguishing terms, generating and binding the candidate set summary and the candidate request summary, executing eight deterministic checks, idempotent deduplication of the source key, and enabling, hierarchical rollback, and disabling the relation calibration group. The centralization and bounded reconstruction module is responsible for determining the cumulative observation mask and complete relation rows, statistically analyzing and branching the common exclusion index and the dispersed exclusion index, determining the lower bound of the three relation reliability factors, centralization by complete relation rows, and projection and assertion of the two channels. The monotonic rearrangement and output module is responsible for re-scoring, shared monotonic calibration, determining the final submission condition, and writing the results.

[0253] In one implementation, the system further includes an auditing and transaction module for storing the pre-update checkpoints, performing atomic commits and overall rollbacks, maintaining an independent monotonic resource ledger, and recording the pre-update state version, temporary state, and ledger version for offline traceability. Optionally, the above modules can be further divided into sub-modules such as input normalization and target anchoring, candidate recall and summary generation, competition analysis and relation domain management, model interface, relation calibration gating, candidate relative feedback calculation, dual-channel projection, atomic checkpoints, monotonic resource ledger, and auditing. Each sub-module can be deployed on the same computing device or collaboratively deployed via a controlled network.

[0254] The present invention also provides an electronic device, including a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the medical terminology standard concept mapping method described in any of the foregoing embodiments. The electronic device may further include a communication interface and a bus. The communication interface is used to access an electronic medical document source system, a standard terminology database service, and the inference service of the generative language model. The processor, the memory, and the communication interface communicate with each other via the bus. The processor may be a central processing unit, an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention. The memory may include random access memory (RAM) or non-volatile memory.

[0255] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when processed, implements the medical terminology standard concept mapping method described in any of the foregoing embodiments. The computer-readable storage medium includes volatile or non-volatile, removable or non-removable media implemented by any method or technology for storing information, including but not limited to random access memory, read-only memory, erasable programmable read-only memory, flash memory, optical disc read-only memory, digital multifunction optical disc, magnetic tape cassette, disk storage, or other magnetic storage devices. The present invention also provides a computer program product comprising a computer program, wherein the computer program, when processed, implements the method described in any of the foregoing embodiments.

[0256] Example 14

[0257] This embodiment provides a deployment architecture, data structure, and degradation handling implementation method for the aforementioned embodiments.

[0258] In one implementation, the system is divided into an offline construction subsystem and an online mapping subsystem. The offline subsystem is responsible for constructing the standard terminology database and candidate signatures, generating the retrieval index and term statistics, training the fixed scoring function, training the two-level probability calibration, defining the closed-set hint template and output mode, and estimating and releasing the conditional confusion matrix. The online subsystem is responsible for target anchoring, candidate recall, closed-set identification, bounded reconstruction, submission decision-making, and auditing. The offline version package is bound to the standard terminology database and signatures, retrieval index and term statistics, the fixed scoring function, the two-level probability calibration, the closed-set hints and output mode, the model and the conditional confusion matrix, as well as submission, rollback, and budget thresholds. Different version packages cannot be used for the same request.

[0259] The main data objects in the online portion include: entity mention objects with original text span; mapping unit objects containing unique targets and binding contexts; term objects storing normalized values, spans, entity internal / external positions, and medical types; candidate concept objects; mapping request objects binding candidate generations, relation domains, candidate set summaries, candidate request summaries, and transaction fields; relation entry objects containing original relations, calibrated relations, source keys, and gating results; iterative state objects serving as rollbackable mapping data states, storing candidate generations, relation domains, feedback rounds, fixed basic matching matrices, shared term weights, candidate residuals, two-level probabilities, and the accumulated decentralized original relations within the current relation domain; resource ledger objects independent of the mapping data state and only monotonically advancing; and mapping response objects binding target identifiers, matching states, standard concept identifiers, probabilities, versions, and audit identifiers. Processing states are distinguished as normal, degraded, and error; mapping states are distinguished as matched, unmatched, and requiring review; infrastructure failures must not be represented as unmatched.

[0260] Furthermore, the status table adopts an append-only or version condition update method and supports atomic acceptance and rollback; the monotonic resource ledger is atomically deducted before external calls or retries and is not restored during rollback; checkpoints are encrypted and saved in their original form, or only the candidate set summary and the candidate request summary, necessary spans, and version information required for auditing are saved.

[0261] It is worth noting that this embodiment handles anomalies and degradation deterministically. The transformed view obtained from optical character recognition only generates a costly retrieval view without overwriting the original text; multiple expanded forms of abbreviations are recalled by versioned indexes; dynamic updates are not performed when there are view conflicts, insufficient context, or low target boundary quality. Temporary states are discarded when model calls time out, there are structural errors, versions are not registered, calibration tables are missing, non-finite values ​​appear, out-of-bounds errors occur, or budget is exhausted; syntax errors allow one retry to fix only the format, and subsequent model calls are stopped when members continuously go out of bounds. Late responses are blocked by random numbers and data state versions, and no degradation path modifies the standard terminology library or global term weights.

[0262] It should be noted that caching and privacy handling are also subject to the aforementioned binding relationships. A summary can be generated from the complete request payload after standard sorting, and a cache key can be constructed using the tenant key and a hash-based message authentication code. The cache must not be reused if the target, context, term span, candidates and their order, relation set, or any version changes. Only verified relations are cached; global term weights are not written back. Remote calls only send de-identified target mentions, necessary binding context, and the simplest description distinguishing terms from candidates, and implement access control and encryption for transmission, storage, and audit logs. The system does not send complete electronic medical documents, target object identifiers, or content unrelated to the current mapping unit to the generative language model. Data processing follows the principles of de-identification and minimum necessity, and is preferably deployed in an institutional intranet or private network.

[0263] Example 15

[0264] This embodiment provides the method for determining parameters, the method for verification and registration, and the analysis of computational complexity in the foregoing embodiments.

[0265] The parameters of the fixed scoring function and their monotonic constraints, coverage calibration and the shared monotonic calibration, calibration temperature, multiplier box factor of shared term weights, update step size of the two channels, total budget of the two classes, relation threshold, first threshold and second threshold, and commit and rollback thresholds are all trained on the development set, determined on an independent calibration set, and evaluated on a locked test set. The conditional confusion matrix is ​​estimated using a separate term-candidate relation annotation set and must not replace the relation truth value with the final mapped label; otherwise, the relation calibration will inherit the label bias of the ranking task.

[0266] Specifically, each learning component is constructed as follows. The fixed scoring function takes the candidate basic feature vector and the candidate directional feedback feature as input, and uses pairwise ranking loss or negative log-likelihood as the training objective on the mapping samples containing correct concept annotations. On the independent development set, the model is selected using first-order accuracy and ranking metrics. In the linear implementation, the coefficients of the candidate directional feedback feature are strictly positive; in the nonlinear implementation, a monotonic shape constraint is applied to the feature, and a monotonicity test is performed before deployment. The construction methods for coverage calibration and the shared monotonic calibration are as described in Example 11. The conditional confusion matrix is ​​obtained by the annotators from the validation set of independently annotated term-candidate relationships, using additive smoothing, and is enabled, hierarchically backed up, or disabled based on the lower limit of the grouped sample size and the lower bound of the confidence interval of the registration performance metrics.

[0267] In one implementation, the following non-limited example parameters can be used for the recalculation of algorithm finiteness and parameter domain checks: maximum feedback rounds are three, supplementary recall budget is one, the upper limit of the number of competing candidates is four, the upper limit of the number of distinguishable terms is three, the lower bound factor of the shared term weight multiplier is 0.5, the upper bound factor is two, the total budget of the shared term weight in norm 1 is 0.4, the absolute value boundary of the candidate residuals is 0.5, the total budget of the candidate residuals in norm 1 is 1.2, the update step size of both channels is 0.5, and the calibration temperature is one. The parameter domain must at least satisfy: the recall channel coefficient and the relation source weight are non-negative; the coefficients of the candidate directional feedback features, the calibration temperature, and the update step size of the two channels are strictly positive; the lower bound factor of the shared term weight is greater than zero and not greater than one, and the upper bound factor is not less than one; the absolute value boundary of the total budget of the two classes and the candidate residuals is non-negative; each probability threshold and the effective relation ratio threshold are between zero and one; and each budget is a non-negative integer. It should be understood that the above values ​​are for illustrative purposes only, and the actual deployment values, projection tolerance, and maximum number of iterations must be determined, registered, and locked on an independent dataset.

[0268] Furthermore, the validation metrics include at least the first-order precision, the top-several-order recall, the macro average F1 score, the false submission rate, the unmapped recall rate, the Brier score or negative log-likelihood, the closed set method rate, the original response member out-of-bounds rate, the proportion of illegal responses entering the accepted state, the interception rate for double summary and relation domain mismatch, the missing item interception rate, the detection rate and false alarm rate of the coverage alarm, the reordering improvement rate and rollback rate, the group deletion sensitivity, the score monotonicity violation rate on the candidate directional feedback feature, the candidate order reversal rate after calibration, the candidate oscillation amplitude, the cross-round drift, and the 95th percentile latency and resource cost, and are statistically segmented by location, document type, source institution, word frequency, and target boundary quality. Ablation groups should include at least: static weighted baselines; direct candidate selection by the model; direct superposition of model scores; removal of matrix integrity requirements; removal of probability centering; use only shared term weights without candidate residuals; removal of boundary constraints; removal of sign preservation constraints; no constraints on the monotonicity of scoring and calibration; detection of only the common exclusion indices and not the dispersed exclusion indices; no distinction between two-level probabilities; use of a single summary or no isolation of relation domains; removal of de-group stability and atomic rollback; and a complete scheme. It should be noted that when forming any performance conclusions, the verification record should include the actual measured values, sample size, fragment definition, and confidence intervals. Indicators that have not been measured should not be written as achieved technical effects. Fragments that do not meet the registration performance or security thresholds should be returned to the baseline through relation calibration gating.

[0269] The computational complexity and resource feasibility are as follows. The target unit has upper limits on the number of terms, recall channels, number of results returned per channel, actual number of candidates, upper limit on the number of aliases per candidate, number of competing candidates, number of distinguishing terms, and number of independent feedback groups. When using hash deduplication and bounded candidate selection, the expected time for candidate fusion is related to the product of the number of channels and the number of results per channel; the time for constructing a fixed term-candidate matching matrix and basic scoring is related to the product of the number of terms, number of candidates, and upper limit on the number of aliases; the payload size of a single template request is related to the number of distinguishing terms, number of competing candidates, and their respective length upper limits, where the size of the closed-set relation grid itself is the product of the number of distinguishing terms and number of competing candidates, and a fixed number of equivalent templates only introduce a fixed multiple. When performing a complete reconstruction from the initial anchor point, in addition to the fixed basic matrix, it is necessary to scan the accumulated original observations in the current relation domain and complete centering, two projections, scoring, and calibration. The time required is related to the number of observations and the worst-case overhead of registering one projection. It is worth noting that the group deletion sensitivity test cannot be simplified to a single reconstruction: there are at most deletion cases with the same number of independent groups. In each case, centering, projection, scoring, and calibration must be re-executed from the candidate generation parameter anchor point while keeping the snapshot unchanged. Therefore, the conservative time for all group deletion checks is related to the square of the number of independent groups and the cost of a single reconstruction, while the workspace does not increase accordingly. This process neither re-selects competing candidates or distinguishes terms, nor calls the generative language model again. Since the scale and number of iterations are all capped, the time and memory within a single request have definite upper bounds; signature pre-computation, relational grid batch processing, and multi-channel parallelism can further reduce online latency, while requests that meet the direct submission conditions do not call the model at all.

[0270] Example 16

[0271] This embodiment illustrates the alternative implementation methods and applicable boundaries of the foregoing embodiments.

[0272] The enumerated relations can take the form of three-valued, five-valued, or discrete intervals, but they must all be mapped to bounded values ​​and calibrated by the conditional confusion matrix before entering parameter updates. The generative language model can be deployed locally, remotely under control, or through controlled multiple voting; regardless of the deployment method, the core feedback path still requires the model to output only fixed terms-candidate relations, and the final mapping result is submitted by the deterministic module. Optionally, a non-generative relation classifier can be used as a degradation baseline to replace the relation generation stage when the model is unavailable or privacy policies prohibit its use; its output must also undergo relation calibration and full deterministic verification.

[0273] Furthermore, the fixed scoring function can adopt an interpretable tree model, a generalized additive model, or a lightweight neural ranking model that satisfies monotonic constraints, but the initial score and the re-scoring should be bound to the same version; the shared monotonic calibration should adopt a transformation that is shared across all candidates, strictly incremental, and equally variable to the candidate arrangement, in order to maintain the consistency between the candidate score order and the probability order.

[0274] It should be noted that the scope of application of this invention is clearly defined. When a standard terminology database provides a unique and accurate mapping, there is no substantial competition between candidates, or the input text significantly exceeds the coverage of the terminology database, the closed-set identification step can be skipped, and the text can be directly submitted or output as an unmapped state. This invention is particularly applicable to situations where a limited number of candidates are highly similar to each other, and there are distinguishing terms in the original text that can be bound across a span. It should be understood that the basic combination advocated by this invention is: term selection based on the same pre-update candidate conditional probability snapshot and signature mutual information with a fixed basic matching strength; dense closed-set relations bound by order-independent candidate set summaries and order-related candidate request summaries respectively; relation domain isolation and machine integrity verification; probability centering; symbol-preserving dual-channel anchoring projection; a fixed scoring function with monotonically unchanging candidate directional feedback features and rank-preserving calibration; and read-only constraints when coverage probability isolation and candidate set summaries remain unchanged. The preferred implementation is a secure closed loop consisting of the pre-update checkpoint, candidate generation, rollbackable mapping data state, independent monotonic resource ledger, and atomic rollback. The recall fusion method, equivalent template form, caching strategy and projection solver can all be replaced, but the above data dependencies and invariants must not be changed.

[0275] The above embodiments are used to illustrate the technical principles of the present invention and do not constitute a limitation on the scope of protection. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for mapping standard concepts in medical terminology, characterized in that, include: Construct mapping units based on target references in medical texts, and recall candidate sets from a standard terminology database; The term-weighted sum of shared term weights and initial zero candidate residuals is used as the candidate directed feedback feature. The candidate score is obtained by a fixed scoring function and calibrated as the candidate conditional probability. The candidate set coverage probability is calculated only using the recall coverage feature before feedback of closed set relation. The candidate conditional probabilities before the update are frozen as a snapshot, based on which the competing candidate set and distinguishing terms are determined. After a closed set identification request, the generative language model returns only the enumeration relationship of the distinguishing terms to the competing candidate set, and after integrity verification, it is converted into bounded signed relation values. The complete relation is centered according to the normalized probability of the snapshot in the competing candidate set to obtain the candidate relative feedback with a weighted sum of zero, based on which the shared term weights and sign-preserving candidate residuals are reconstructed from the initial values ​​within the item-by-item boundary and the total budget. By re-scoring the candidate directional feedback features using the fixed scoring function that does not decrease monotonically, updating the candidate conditional probabilities through shared monotonic calibration, and outputting the standard conceptual identifier of the first candidate, which is then written into the medical data record.

2. The method according to claim 1, characterized in that, The step of freezing the candidate conditional probabilities before the update as a snapshot, and determining the competing candidate set and distinguishing terms, includes: The basic matching strength of each candidate in the candidate set on each term is discretized into signature bins according to the registration threshold; The signature information gain of each term is calculated using the normalized probability of the snapshot within the competing candidate set and the signature binning. By using a pre-registered combination function, the query value of each term is determined by taking the signature information gain, the term being located within the target mention, the term's medical type, the term's negation domain, and the term's span quality as gain terms, and the term's redundancy relative to the selected terms and the term's sensitive information transmission cost as subtraction terms. Select up to a preset number of terms in descending order of query value to form the distinguishing terms; If the query value is lower than the registered value threshold or there are fewer than two candidates in the competitive candidate set, the closed set identification request will not be sent, and the process will proceed to supplementary recall or output a manual review mark.

3. The method according to claim 2, characterized in that, The step of calculating the signature information gain of each term using the normalized probability of the snapshot within the competing candidate set and the signature binning includes: The normalized probability of the snapshot within the competing candidate set is used as the distribution of the candidate random variable; The signature bins of each competing candidate on the same term are used as the signature random variables for that term; The signature information gain is defined as the difference between the entropy of the candidate random variable and the conditional entropy of the candidate random variable under the signature random variable condition. The conditional entropy is calculated by grouping the competing candidate set according to the signature bins, and the signature bins with a probability mass of zero are not included in the summation of the conditional entropy; When all candidates in the competitive candidate set have the same signature binning on the same term, the signature information gain of that term is zero. When the numerical error causes the signature information gain to be negative, the signature information gain is clipped to zero.

4. The method according to claim 1, characterized in that, The closed-set identification request causes the generative language model to return only the enumeration relation of the distinguishing terms to the competing candidate set, which is then converted into bounded signed relation values ​​after integrity verification, including: Based on the standard terminology database version and the candidate set ordered according to the immutable candidate identifier specification, an order-independent candidate set summary is generated; Based on the span mentioned in the target, the distinguishing terms and the span of the distinguishing terms, the set of competing candidates arranged in the order of this presentation, and the version information, a sequence-related candidate request summary is generated; The closed set identification request and the response to the closed set identification request are simultaneously bound to the candidate set summary, the candidate request summary and the relation domain identifier, and the values ​​of the enumerated relations are limited to three enumerated values: support, exclusion and irrelevant. The response is evaluated for format, request binding, membership, uniqueness, and privilege escalation. If the response is missing any of the requested terms / candidate pairs, the entire response is deemed invalid. For the same term-candidate pair, at least two equivalent requests with different candidate presentation orders are used. Each equivalent request is bound to a corresponding candidate request summary. When the enumeration relations returned by the equivalent requests are inconsistent, the term-candidate pair is judged as unobserved, and the relation row containing the term-candidate pair is not included in the subsequent centralization.

5. The method according to claim 4, characterized in that, The conversion of the integrity-verified value into a bounded, signed relational value includes: A relation validation set, which is separate from the candidate ranking training set and has manually labeled terms and candidate relations, is used. Create relation calibration groups based on one or more of the following: model version, prompt template, medical field, relation type, and reason code; Based on the conditional confusion matrix of the enumerated relation output by the generative language model in the relation calibration group relative to the manually labeled relation, the enumerated relation is converted into a signed expectation, which is used as the bounded signed relation value. When the sample size of the relationship calibration group reaches the registered minimum sample size, and the lower bound of the confidence interval of the registered relationship calibration performance index reaches the registered activation threshold, the relationship calibration group is activated; otherwise, it is deactivated by falling back according to the pre-registered reason code, medical field and prompt template level. The transformation does not use the confidence level reported by the generative language model and does not combine the statistical data of the conditional confusion matrix across model versions without independent verification.

6. The method according to claim 1, characterized in that, The method further includes: Before the reconstruction of the centralization, the shared term weights, and the candidate residuals, the weighted proportion of complete relation rows in the competitive candidate set that are mutually exclusive is calculated using the initial shared term weights and the relation reliability independent of the sign of the relation value as statistical weights, to obtain the common exclusion index. The weighted proportion of complete relation rows that exclude a candidate is calculated separately for each candidate with the same statistical weight, and the minimum value is taken for all candidates in the competitive candidate set to obtain the dispersed exclusion index; When the common exclusion index reaches the first threshold or the dispersed exclusion index reaches the second threshold, a coverage alarm is triggered. When the coverage alarm is triggered, the bounded signed relation values ​​of this batch are written only to the audit record, and the shared term weight, the candidate residual, the candidate score, the candidate conditional probability and the candidate set coverage probability are kept to their previous values. When the coverage alert is triggered and there are candidates in the candidate set that have not entered the competing candidate set, the competing candidate set is expanded and a new relational domain is established; otherwise, supplementary recall is performed when the supplementary recall budget allows. If the supplementary recall budget is insufficient, an unmapped status or a manual review mark is output.

7. The method according to claim 6, characterized in that, The statistical weights, which are the initial shared term weights and the relation reliability independent of the sign of the relation value, include: For each complete relation row, determine the relation calibration reliability factor, source reliability factor, and structural verification reliability factor respectively; Each of the reliability factors is given by a registration function bound to the model version, hint version, relation calibration version, and structure verification rule version, and takes a value between zero and one, independent of the sign of the bounded signed relation value; The minimum value among the relationship calibration reliability factor, the source reliability factor, and the structural verification reliability factor is taken as the relationship reliability. The statistical weight is the product of the initial shared term weight of the term corresponding to the complete relation row and the relation reliability. When the sum of the statistical weights of all complete relation rows is zero, the closed set relation feedback in this round is deemed invalid, and the candidate condition probability and the candidate set coverage probability are kept at their previous values.

8. The method according to claim 1, characterized in that, The step of centering the complete relation according to the normalized probability within the competitive candidate set based on the snapshot, to obtain candidate relative feedback with a weighted sum of zero, includes: For the bounded signed relation values ​​accumulated up to this round within the current relation domain, determine the cumulative observation mask by term-candidate pair; When all candidates in the competing candidate set have been observed for the same term, the relation row corresponding to the term is determined to be a complete relation row. Relation rows with unobserved term-candidate pairs are set to zero and do not participate in the centralization. For each complete relation row, the bounded signed relation value of the complete relation row is weighted by the normalized probability of the snapshot in the competing candidate set, and the bounded signed relation value of each candidate in the complete relation row is subtracted from the weighted average to obtain the candidate relative feedback. When the bounded signed relation values ​​of all candidates in the competing candidate set are the same for the same complete relation row, the relative feedback of the candidate in that complete relation row is zero. Each feedback round uses the snapshot described in this round to uniformly re-centralize all complete relations accumulated in the current relation domain, without centralizing the already centralized results again; When the members of the competing candidate set change, a new relation domain is established, the feedback rounds in the new relation domain are reset to zero, and the bounded signed relation values ​​of the previous relation domain do not participate in the centralization in the new relation domain.

9. The method according to claim 1, characterized in that, The method for reconstructing the shared term weights and sign-preserving candidate residuals from initial values ​​within the term-by-term boundaries and the total budget includes: The distinguishing strength of each term is determined by the weighted L2 norm of the normalized probability of the snapshot within the competing candidate set, based on the relative feedback of the candidates. The initial shared term weights are multiplicatively updated and normalized according to the distinguishing strength, and then projected onto the feasible region that simultaneously satisfies the weight sum of one, the multiplier box constraint relative to the initial shared term weights item by item, and the first norm total budget constraint, to obtain the reconstructed shared term weights. The product of the candidate relative feedback, the non-negative update step size, and the relation validity gating value is used as the candidate residual target value; The candidate residual target value is projected onto the feasible region that simultaneously satisfies the terms-by-term absolute value boundary, the norm total budget, the residual terms not belonging to the complete relation row being zero, the residuals within the complete relation row having a zero mean with respect to the normalized probability weight, and the sign preservation constraint, to obtain the reconstructed candidate residual. The sign-preserving constraint is that the product of the reconstructed candidate residual and the corresponding candidate relative feedback is non-negative, and the candidate residual is zero when the candidate relative feedback is zero. If any of the assertions in the following categories—the absolute value boundary, the norm total budget, the weighted zero mean, and the sign preservation constraint—fail after projection, the current update will not be committed and the process will be rolled back to the checkpoint before the update.

10. The method according to claim 9, characterized in that, The process of projecting the candidate residual target value onto a feasible region that simultaneously satisfies the following constraints: term-by-term absolute value boundary, norm 1 total budget, zero residual terms not belonging to the complete relation row, zero mean of residuals within the complete relation row with respect to the normalized probability weighted average, and sign preservation constraint, yields the reconstructed candidate residual, including: The reconstructed candidate residuals are represented by the product of the coefficient of the same nonnegative row and the candidate relative feedback of the complete relation row. The upper limit of the non-negative row coefficient is taken as the minimum value of the ratio of the absolute value boundary of each non-zero candidate in the complete relation row to the absolute value of the candidate relative feedback. Set the coefficients of the non-negative rows that do not belong to the complete relation rows to zero; The sum of the products of the non-negative row coefficients of each complete relation row and the first norm of the candidate relative feedback of that complete relation row shall not exceed the total budget of the first norm; The weighted sum of the candidate relative feedbacks with respect to the normalized probabilities is zero, so that the candidate residuals obtained by scaling with the non-negative row coefficients satisfy the weighted zero mean and the sign retention constraint.

11. The method according to claim 4, characterized in that, The process of obtaining candidate scores from a fixed scoring function and calibrating them into candidate conditional probabilities, and calculating the candidate set coverage probability using only the recall and coverage features before feedback of the closed set relation, includes: The recall coverage features include at least one of the following: maximum recall score, candidate coverage rate, number of healthy recall channels, proportion of unregistered medical terms, and span quality and type conflict of the target mention; The recall coverage features do not include the enumeration relationship, the reconstructed shared term weights, the candidate residuals, and the candidate targeted feedback features; The candidate set coverage probability is subtracted from the candidate set coverage probability to obtain the candidate set out-of-set probability, and the joint probability of each candidate is obtained by multiplying the candidate set coverage probability and the candidate conditional probability. The calibration of the candidate conditional probability and the calculation of the candidate set coverage probability use mutually isolated data paths; When the candidate set summary remains unchanged, the candidate set coverage probability and the candidate set out-of-set probability are kept as read-only. When the candidate set summary changes due to the supplementary recall, a new candidate generation is established, the candidate set summary is regenerated, and the candidate set coverage probability is recalculated with the new recall coverage features.

12. The method according to claim 1, characterized in that, The method further includes: At the beginning of each feedback round, the current accepted mapping data state is saved as a checkpoint before update. The mapping data state includes candidate generation, relation domain, candidate set, competing candidate set, initial shared term weight, shared term weight, candidate residual, candidate directional feedback feature, candidate score, candidate conditional probability, candidate set coverage probability, and the cumulative uncentralized relation and data state version in the current relation domain. If the request is consistent with the data state version, and the relation integrity, numerical finiteness, probability isolation, item-by-item boundary, total budget, sign preservation and weighted zero mean assertions all pass, the updated mapped data state is atomically committed; otherwise, the entire process is rolled back to the checkpoint before the update. The number of feedback rounds, the number of replenishment recalls, and the number of format repairs are maintained by a monotonic resource ledger independent of the state of the mapping data. These resources are atomically deducted before each call to the generative language model, and the deducted resources are not restored when the state of the mapping data is rolled back. The standard conceptual identifier of the first candidate is written into the medical data record on the premise of meeting the final submission conditions. The final submission conditions include the joint probability of the first candidate reaching the submission threshold, the joint probability interval between the first candidate and the second candidate reaching the interval threshold, the out-of-set probability of the candidate not exceeding the out-of-set threshold, the effective relationship ratio reaching the relationship threshold, and the first candidate belonging to the completely identified set of competing candidates. When the processing budget is exhausted, the deadline is reached, or an irrecoverable verification failure occurs and the final submission condition is not met, an unmapped status or a manual review mark is output.

13. A standard concept mapping system for medical terminology, characterized in that, include: The mapping unit construction module is used to construct mapping units based on target references in medical texts and to recall a candidate set from a standard terminology database. The initial scoring and two-level probability module is used to obtain candidate scores and calibrate them into candidate conditional probabilities by using the term-weighted sum of shared term weights and candidate residuals that are initially zero as candidate-directed feedback features. The candidate scores are obtained by a fixed scoring function and calibrated into candidate conditional probabilities. The candidate set coverage probability is calculated only using the recall coverage features before feedback of closed set relations. The closed set identification and verification module is used to freeze the candidate conditional probabilities before the update as a snapshot, thereby determining the competing candidate set and distinguishing terms. After the closed set identification request, the generative language model returns only the enumeration relationship of the distinguishing terms to the competing candidate set, and after the integrity verification, it is converted into a bounded signed relation value. The centralization and bounded reconstruction module is used to centralize the complete relation according to the normalized probability of the snapshot in the competing candidate set, to obtain the candidate relative feedback with a weighted sum of zero, so as to reconstruct the shared term weights and sign-preserving candidate residuals from the initial values ​​within the item-by-item boundary and the total budget; The monotonic rearrangement and output module is used to re-evaluate the candidate directional feedback features by the fixed scoring function that is monotonically undiminished, update the candidate conditional probabilities through shared monotonic calibration, and output the standard concept identifier of the first candidate to be written into the medical data record.

Images