Multi-source heterogeneous long text double-track entity recognition spoken language denoising analysis method and system

Abstract

The application relates to the technical field of data processing, and discloses a double-track entity recognition spoken language denoising analysis method and system for multi-source heterogeneous long texts. The method comprises the following steps: performing double-layer disassembly on original daily conversation texts of patients to generate a clause-level semantic unit set with time stamps and serial number, constructing a two-dimensional speech flow conversion space with an ontology feature, an object interaction and an environmental noise flow coordinate layer as longitudinal axes, and obtaining the above space with cross-layer dependency pointers; identifying a shielding time window according to a medical entity density parameter in the object interaction flow coordinate layer to generate a shielding protection buffer distance interval; performing subject-predicate-object relationship re-inspection processing on medical entities in the ontology feature flow coordinate layer in the interval to generate a structured case closing summary text; and the application adopts a multi-subject space isolation and conditional shielding strategy, avoids subject confusion and medical feature misattribution in a complex context, and guarantees the attribution accuracy of medical data.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically, to a dual-track entity recognition spoken language denoising and parsing method and system for multi-source heterogeneous long text. Background Technology

[0002] With the application of medical informatization in the field of smart healthcare, extracting structured data from massive amounts of doctor-patient dialogue texts has become a crucial step in generating electronic medical records and case summaries. Patent CN121303334B discloses a method for understanding and generating medical dialogues based on the Transformer model. This technology uses a Perceiver structure to fuse medical corpora across sources and combines neural network differential equations to dynamically model the semantic evolution process between dialogue turns to guide text generation, ensuring consistency in semantic trajectory and entity reference in the generated medical dialogue responses. Patent CN120893435B discloses a method and system for text entity recognition and data structuring. This technology uses a two-stage entity recognition model to extract indicator items from medical diagnostic texts and transform them into structured entities. It also incorporates closed-loop correction based on sampling verification results from medical personnel. Through a combination of model recognition and manual verification, it ensures the accuracy and completeness of the final generated structured medical data.

[0003] However, in actual outpatient follow-up medical record archiving scenarios, when the system processes long asynchronous text messages between doctors and patients that include everyday details, it is prone to incorrectly including medications and symptoms of the patient's family members or pets in the patient's own case summary. For example, after a patient states that they are taking their hypoglycemic medication on time, and then adds information about their family member having a fever, taking antipyretics, and experiencing diarrhea, the system's extracted "Recently Added Medications and Adverse Reactions" results may include medical entities such as "antipyretics" and "diarrhea" belonging to the family member. This crosstalk in feature extraction caused by multiple subjects' overlapping narratives leads to the extracted medical labels being detached from the patient's actual condition. Once medical indication data containing incorrect subjects is aggregated and entered into the discharge summary, it will cause subsequent follow-up doctors to make misjudgments based on the initial baseline data with attribution bias, thus providing incorrect medication dosage adjustments or contraindications for intervention. Summary of the Invention

[0004] The misattribution of medical entities in the generated structured case summary stems from a structural misalignment between the linear temporal arrangement of asynchronous communication text and the interweaving of multidimensional logical subjects. The underlying large language model primarily relies on a self-attention mechanism when handling extremely long contexts. This mechanism tends to capture medical and symptom entities that are close together within the same analysis window during computation. When everyday spoken language interactions involve cross-subjective narratives across different scenarios, the distance weight of positional encoding in continuous text sequences outweighs the dependency weight of logical subjects. The parsing engine lacks a robust anti-crosstalk barrier mechanism to physically isolate interwoven threads during feature matching, failing to effectively remove subject crosstalk caused by cross-narratives. This leads to semantic collapse of the subject and object during noise reduction parsing, ultimately resulting in incorrect attribution of extracted medical features due to feature matching crossover and misattribution of medical labels.

[0005] To overcome the shortcomings of existing technologies, such as crosstalk and attribution errors in medical entity feature extraction due to multi-subject cross-narrative, this invention provides a dual-track entity recognition and speech denoising parsing method and system for multi-source heterogeneous long texts. This method involves decomposing the original text of patients' daily conversations into two layers and constructing a two-dimensional discourse flow space with different attribution layers along the vertical axis. Based on the density parameter of medical entities within the object interaction flow coordinate layer, a shielding buffer distance interval is dynamically generated. Within this interval, the subject-verb-object relationship of medical entities within the ontology feature flow coordinate layer is re-verified. This invention constructs a conditional shielding and boundary parameter adaptive scaling mechanism based on the subject space state, eliminating subject referential confusion in complex spoken interactions of multi-source long texts, and ensuring the consistency and logical rigor of the extracted and aggregated data with the baseline data of the discharge summary.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A dual-track entity recognition spoken language denoising and parsing method for multi-source heterogeneous long texts includes: The system receives the original text of patients' daily conversation records, performs a two-layer decomposition on the original text, generates a set of clause-level semantic units with timestamp tags and ordinal numbers, and constructs a two-dimensional discourse flow space. The two-dimensional discourse flow space is divided into an ontology feature flow coordinate layer, an object interaction flow coordinate layer, and an environmental noise flow coordinate layer on the vertical axis. Based on the set of clause-level semantic units with timestamp tags and ordinal numbers and the two-dimensional discourse flow space, a two-dimensional discourse flow space with cross-layer dependency pointers is obtained. Based on a two-dimensional discourse flow space with cross-layer dependency pointers, the density parameters of medical entities in the object interaction flow coordinate layer of the two-dimensional discourse flow space are calculated. Based on the density parameters of medical entities, the time windows that need to be blocked are identified, and a blocking protection buffer distance interval is generated for each time window that needs to be blocked. Within the blocking protection buffer distance interval, the subject-verb-object relationship re-verification processing is performed on the medical entities in the ontology feature flow coordinate layer, and a structured case summary text is generated.

[0007] A dual-track entity recognition spoken language denoising and parsing system for multi-source heterogeneous long texts, used to implement the aforementioned dual-track entity recognition spoken language denoising and parsing method for multi-source heterogeneous long texts, the system comprising: The decomposition space construction module is used to receive the original text of the patient's daily dialogue record and the patient's discharge summary document, perform two-layer decomposition on the original text of the patient's daily dialogue record, generate a set of clause-level semantic units with timestamp tags and sequence numbers, and construct a two-dimensional discourse flow space. The two-dimensional discourse flow space is divided into an ontology feature flow coordinate layer, an object interaction flow coordinate layer and an environmental noise flow coordinate layer on the vertical axis. Cross-layer pointer construction module: Based on the set of clause-level semantic units with timestamp tags and ordinal numbers and the two-dimensional discourse flow space, a two-dimensional discourse flow space with cross-layer dependency pointers is obtained; The shielding interval generation module: Based on the two-dimensional discourse flow space with cross-layer dependency pointers, it calculates the density parameters of medical entities in the object interaction flow coordinate layer of the two-dimensional discourse flow space, identifies the time windows that need to be shielded based on the density parameters of medical entities, and generates the shielding protection buffer distance interval for each time window that needs to be shielded. Entity Verification and Closure Module: This module is used to perform subject-verb-object re-verification processing on medical entities within the ontology feature flow coordinate layer within the shielded protection buffer distance range, and generate a structured closure summary text.

[0008] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention deconstructs the original text of patients' daily conversations into a two-dimensional discourse flow space by dividing the vertical axis into an ontology feature flow coordinate layer, an object interaction flow coordinate layer, and an environmental noise flow coordinate layer. This transforms the originally linear, unstructured, interwoven text into a parallel spatial structure with a state of belonging and isolation, allowing the semantic units of different clauses to be physically separated and preventing crosstalk in feature extraction under multi-subject cross-narrative conditions. Furthermore, based on the density parameter of medical entities within the object interaction flow coordinate layer, the invention identifies the time window that needs to be masked and generates a masking protection buffer distance interval. This constructs a dynamic masking boundary that depends on the frequency of occurrence of medical topics. Within this buffer interval, subject-verb-object relationship re-verification processing is specifically initiated for medical entities within the ontology feature flow coordinate layer. This solution can target and intercept high-probability subject confusion intervals in complex everyday spoken language interactions, eliminating the misattribution of medical events and misclassification of features caused by frequent switching of the subject in the dialogue (such as the patient's description of their own symptoms being intertwined with the family's condition). It ensures that the verified and finally aggregated medical entities strictly belong to the patient, and that the medical indicator data in the generated structured case summary text has objective authenticity and accurate subject correspondence. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a flowchart of the dual-track entity recognition and speech denoising parsing method for multi-source heterogeneous long text in this invention. Figure 2 This is a schematic diagram of the double-layer disassembly process in this invention; Figure 3 This is a schematic diagram of the two-dimensional discourse flow spatial layering in this invention; Figure 4 This is a schematic diagram of baseline anchoring mapping in this invention; Figure 5 This is a schematic diagram illustrating the density of medical entities within an equal-width time window in this invention. Figure 6 This is a schematic diagram of the shielding protection buffer distance range in this invention; Figure 7 This is a flowchart illustrating the verification process for the attribution of medical entities within the shielding and protection buffer distance range in this invention. Figure 8This is a functional block diagram of the dual-track entity recognition and speech denoising parsing system for multi-source heterogeneous long texts in this invention. Detailed Implementation

[0011] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0012] Example 1: Please see Figure 1 As shown, this embodiment provides a dual-track entity recognition spoken language denoising and parsing method for multi-source heterogeneous long texts, including: Step S10: Receive the original text of the patient's daily conversation records and the patient's discharge summary document. Perform a two-layer decomposition on the original text of the patient's daily conversation records to generate a set of clause-level semantic units with timestamp tags and sequence numbers, and construct a two-dimensional discourse flow space. The two-dimensional discourse flow space is divided into an ontology feature flow coordinate layer, an object interaction flow coordinate layer, and an environmental noise flow coordinate layer on the vertical axis. Extract the discharge summary structured baseline dataset from the patient's discharge summary document. Based on the discharge summary structured baseline dataset and the two-dimensional discourse flow space, obtain the two-dimensional discourse flow space with cross-layer dependency pointers and the ontology feature flow coordinate layer data with baseline anchoring mapping.

[0013] Specifically, step S10 aims to integrate two types of long text data—the original text of the patient's daily conversation records and the patient's discharge summary document—which have different sources and structures, into a unified analytical framework. The original text of the patient's daily conversation records is unstructured spoken text, containing a large amount of noise information such as small talk unrelated to the patient's own condition, paraphrasing of others' symptoms, and environmental chatter. Furthermore, the same message may contain multiple semantic fragments with multiple logical subjects pointing to different objects. The patient's discharge summary document is structured medical text, recording clinically confirmed diagnostic information, test indicators, and treatment plans, possessing high-confidence medical factual baseline characteristics. Step S10 performs a two-layer decomposition process on the original text of the patient's daily conversation records, breaking down the continuous dialogue flow into the smallest independently analyzable semantic particles. Then, based on the logical subject identification results, it projects semantic particles of different affiliations layer by layer in the two-dimensional discourse flow space. This allows the subsequent step S20 to distinguish between patient self-reported information and non-patient information in the spatial dimension, avoiding the misclassification of others' symptoms or environmental noise into the scope of patient entity recognition. Meanwhile, step S10 extracts the structured baseline dataset of the discharge summary from the patient's discharge summary document and establishes an anchoring mapping with the medical entities in the ontology feature flow coordinate layer, providing a clinically validated reference benchmark for entity aggregation and structured case summary text generation in step S20.

[0014] Further, step S10 includes: Step S11: Receive the original text of the patient's daily conversation records and the patient's discharge summary document. Perform a two-layer decomposition process on the original text of the patient's daily conversation records. The two-layer decomposition process includes a first-layer timestamp slicing process and a second-layer clause-level semantic decomposition process. The first-layer timestamp slicing process is: slice the original text of the patient's daily conversation records according to the message sending timestamp to obtain a message slice sequence arranged in chronological order. The second-layer clause-level semantic decomposition process is: for each message slice in the message slice sequence, use a punctuation mark boundary and transition conjunction position joint delimitation mechanism to split it into multiple independent clause-level semantic units; retain the timestamp label and sequence number to construct a set of clause-level semantic units with timestamp label and sequence number.

[0015] See Figure 2 This is a schematic diagram of the double-layer disassembly process provided in the embodiments of this application. Figure 2The diagram illustrates the hierarchical decomposition process of the original text of a patient's daily conversations along the timeline, demonstrating a two-level decomposition process from the original text to message slices and then to sentence-level semantic units, and annotating two types of decomposition delimitation markers. Specifically, the first layer of the two-level decomposition process is timestamp slicing: each message in the original text of the patient's daily conversations carries a message sending timestamp, which records the date and time the patient sent the message. The original text of the patient's daily conversations is sliced ​​according to the order of the message sending timestamps, dividing the original text into segments... Figure 2 The sequence of message slices arranged chronologically is shown, with each timestamp corresponding to the segmentation boundary of the message slice on the timeline. Each message slice in the chronologically arranged sequence corresponds to a complete message sent by the patient at a certain moment. The reason for performing the first-level timestamp slicing processing is that the original text of the patient's daily dialogue records is arranged continuously in the form of a timeline. The topics, symptoms, and medications discussed in messages at different times may change or even contradict each other. If the original text of the patient's daily dialogue records is treated as a whole instead of slicing according to the message sending timestamp, the medical entities in subsequent steps will lose their temporal positioning information, making it impossible to trace the time node when each medical entity appears, and also failing to support the requirement of dividing the time window into equal widths according to the horizontal time dimension and calculating the density parameter of medical entities in step S20.

[0016] The second layer of the two-layer decomposition process is sentence-level semantic decomposition: for each message slice in the chronologically ordered message slice sequence, the following steps are taken: Figure 2 The punctuation mark boundary and adversative conjunction position joint delimitation mechanism, as indicated in the text, decomposes the message into several independent clause-level semantic units, completing the granular refinement from message slices to clause-level semantic units. The operation of the punctuation mark boundary and adversative conjunction position joint delimitation mechanism is as follows: Terminal punctuation marks such as periods, question marks, exclamation marks, and semicolons are used as the first delimitation markers, segmenting the message at the positions of these terminal punctuation marks; within each segmented segment, the position of adversative conjunctions is further detected as the second delimitation marker. Adversative conjunctions refer to connective words in a sentence that indicate a change in semantic direction, including but not limited to words such as "but," "however," "yet," "but," and "indeed," which indicate a semantic contrast, as well as words such as "actually," "anyway," and "just," which are often used in spoken language to introduce supplementary explanations with a different semantic direction than the preceding text. Further segmentation at the positions of adversative conjunctions yields... Figure 2The sentence-level semantic units shown are used. The reason for employing a combined delimitation mechanism of punctuation boundaries and conjunction positions, rather than relying solely on punctuation, is that patients often use commas to connect multiple clauses with different semantic orientations in their spoken expressions. Segmentation based solely on final punctuation would result in a mixture of the patient's own account and what others have relayed within the same segment. Conjunctions often indicate a shift in logical subject or a change in semantic direction. Performing secondary segmentation at the conjunction positions can separate semantic segments pointing to different logical subjects, ensuring that each sentence-level semantic unit contains only a single logical subject, providing a cleaner input particle for logical subject identification in step S12. For example, a message sent by a patient may contain a complex sentence with two different logical subject segments. After final punctuation segmentation, it remains a single segment. After secondary segmentation at the conjunction positions, two sentence-level semantic units are obtained. The former's logical subject points to the patient's own symptom description, while the latter's logical subject points to the health status of the patient's family. These two sentence-level semantic units will be assigned different subject category labels in step S12.

[0017] After completing the semantic decomposition at the clause level, each clause-level semantic unit retains its message sending timestamp as a timestamp label, and assigns an ordinal number according to the order in which the clause-level semantic unit appears in the original message slice, thus constructing a set of clause-level semantic units with timestamp labels and ordinal numbers. The timestamp label is used to record the message sending timestamp of the message slice to which the clause-level semantic unit belongs, and the ordinal number is used to record the positional order of the clause-level semantic unit within the same message slice. The combination of the two allows the clause-level semantic unit to be projected onto a definite coordinate position on the horizontal axis of the two-dimensional discourse flow space in the subsequent step S13. Furthermore, when there are multiple clause-level semantic units under the same message sending timestamp, the ordinal number ensures that the arrangement order among the clause-level semantic units is not lost. Step S11 transforms the unstructured original text of the patient's daily dialogue records into a set of clause-level semantic units with timestamps and sequence numbers through a two-layer decomposition process. This allows subsequent steps to perform logical subject identification and medical entity scanning at the clause-level semantic unit granularity, avoiding entity attribution errors caused by the mixing of multiple logical subjects when processing at the whole message granularity. Without the two-layer decomposition process in step S11, the logical subject identification process in step S12 would face a complete message containing multiple logical subjects. The accuracy of logical subject identification would be reduced due to the mixing of subject references in the input text, which would then affect the correctness of the vertical axis coordinate layer allocation in step S13, ultimately causing the distinction between the ontology feature flow coordinate layer and the object interaction flow coordinate layer in step S20 to lose reliability.

[0018] Step S12: Perform logical subject identification processing on each clause-level semantic unit in the set of clause-level semantic units with timestamp labels and ordinal numbers. Based on the logical subject identification results, assign one of three subject category labels to each clause-level semantic unit: ontology feature flow label, object interaction flow label, or environmental noise flow label, to obtain a set of clause-level semantic units with subject category labels.

[0019] Specifically, in step S12, logical subject identification processing is performed on each clause-level semantic unit in the set of clause-level semantic units with timestamp tags and sequence numbers. Logical subject identification processing refers to semantic analysis of the text content of the clause-level semantic units to determine the subject or sender of the symptoms, feelings, behaviors, or events described in the clause-level semantic units, i.e., the logical subject. Logical subject identification processing employs a subject extraction method based on dependency parsing: word segmentation and part-of-speech tagging are performed on the text of the clause-level semantic units to obtain a part-of-speech tagging sequence; dependency parsing is then performed on the part-of-speech tagging sequence to obtain a dependency parsing tree; word nodes with subject dependency relationships with the predicate verb are extracted from the dependency parsing tree as candidate logical subjects. When the candidate logical subject is a personal pronoun, personal pronoun reference resolution processing is performed in conjunction with the context of the message slice to which the clause-level semantic unit belongs, restoring the personal pronoun to the specific person it refers to, thus determining the referent of the logical subject. The scope of the context includes the entire text content of the message slice to which the clause-level semantic unit belongs, as well as the text content of the two message slices immediately before and after it on the timestamp label. When the text content of the message slice to which the clause-level semantic unit belongs is insufficient to determine the referent of the personal pronoun, the search for the referent is expanded to the adjacent message slices before and after it. If the referent of the personal pronoun still cannot be determined within the context, the logical subject identification result of the clause-level semantic unit is marked as uncertain. When assigning subject category labels, the clause-level semantic unit carrying the uncertain status label is assigned to the environmental noise stream label, so that it does not participate in the subsequent medical entity attribution determination process, avoiding misclassification due to failure of reference resolution. Based on the logical subject identification results, each clause-level semantic unit is assigned one of three subject category labels according to the following classification rules: When the logical subject points to the patient themselves, an ontology feature flow label is assigned, indicating that the clause-level semantic unit describes the patient's own symptoms, feelings, medication feedback, or changes in physiological state; when the logical subject points to someone other than the patient (such as the patient's family members, medical staff, etc.), an object interaction flow label is assigned, indicating that the clause-level semantic unit describes someone else's symptoms, suggestions, or medical-related information, which may contain valid medical entities but should not be attributed to the patient; when the logical subject cannot point to any person or the content of the clause-level semantic unit is unrelated to medical matters, an environmental noise flow label is assigned, indicating that the clause-level semantic unit is non-medical information such as small talk, chatter, or emotional expression.

[0020] After assigning subject category labels, a set of clause-level semantic units with subject category labels is obtained. The logical subject identification process in step S12 enables each clause-level semantic unit to obtain a clear attribution determination, providing a classification basis for projecting clause-level semantic units onto different vertical axis coordinate layers in the two-dimensional discourse flow space in step S13. The reason for using three classification methods—ontology feature flow labels, object interaction flow labels, and environmental noise flow labels—instead of a simple binary classification is that the influence of medical information related to others and pure noise information in the original text of the patient's daily dialogue records on subsequent processing is different: although medical information related to others should not be attributed to the patient, when medical entities related to others appear densely within a certain time window, it may reflect that the patient is recounting their medical experience or discussing medication plans. The patient's self-narrated information is easily mixed with the context of other people's information in the same time interval. Step S20 uses the medical entity density parameter of the object interaction flow coordinate layer to identify such time windows that are prone to attribution confusion; while pure noise information does not contain medical entities and does not interfere with subsequent entity identification. After being classified into the environmental noise flow coordinate layer, it is excluded from the scope of subsequent processing. If medical information related to others is merged with pure noise information into a single non-patient category, step S20 will be unable to extract the temporal distribution of medical information related to others to calculate the medical entity density parameter. The calculation result of the medical entity density parameter will be diluted due to the inclusion of a large number of pure noise clause-level semantic units without medical entities in the object interaction flow coordinate layer, resulting in a decrease in the recognition sensitivity of the time window to be masked. If the logical subject recognition processing and three-category label assignment in step S12 are missing, step S13 will be unable to construct the vertical axis coordinate layer, and the two-dimensional discourse flow space will degenerate into a one-dimensional time series. The mechanism in step S20 that relies on the medical entity density parameter of the object interaction flow coordinate layer to identify the time window to be masked will not function.

[0021] Step S13: Extract the timestamps of the message slice sequence arranged in chronological order as the horizontal axis coordinates, and extract the subject category labels of the sentence-level semantic unit set with subject category labels as the vertical axis coordinate layer to establish a two-dimensional discourse flow space. Project each sentence-level semantic unit in the sentence-level semantic unit set with subject category labels onto the corresponding coordinate position on the corresponding coordinate layer in the two-dimensional discourse flow space according to the timestamp and subject category label. Establish cross-layer dependency pointers for sentence-level semantic units that originate from the same message slice but belong to different coordinate layers, and output the two-dimensional discourse flow space with cross-layer dependency pointers. The coordinate layers include the ontology feature flow coordinate layer, the object interaction flow coordinate layer, and the environmental noise flow coordinate layer.

[0022] See Figure 3 This is a schematic diagram of the two-dimensional discourse flow space layering provided in the embodiments of this application. Figure 3The diagram illustrates a two-dimensional spatial structure with the timestamp as the horizontal axis and the coordinate layer corresponding to the subject category label as the vertical axis. It shows three independent vertical axis coordinate layers and the temporal distribution of clause-level semantic units within each layer. Each circular node represents a clause-level semantic unit. Specifically, in step S13, the message sending timestamp of each message slice in the chronologically ordered message slice sequence is extracted as the horizontal axis coordinate value. The subject category label carried by each clause-level semantic unit in the set of clause-level semantic units with subject category labels is extracted as the vertical axis coordinate layer affiliation. A structure is established with the message sending timestamp as the horizontal axis and the coordinate layer corresponding to the subject category label as the vertical axis. Figure 3 The diagram illustrates a two-dimensional discourse flow space. The vertical axis of this space is divided into three coordinate layers: the ontological feature flow coordinate layer is located in the upper region of the vertical axis, corresponding to… Figure 3 The layered region labeled "Patient's Own" is used to carry clause-level semantic units carrying ontology feature flow labels; the object interaction flow coordinate layer is located in the middle region of the vertical axis, corresponding to... Figure 3 The layered area marked "related to others" is used to carry clause-level semantic units carrying object interaction flow tags; the environmental noise flow coordinate layer is located in the lower layer area of ​​the vertical axis, corresponding to... Figure 3 The layered area labeled "small talk" is used to carry clause-level semantic units carrying environmental noise flow labels. Each clause-level semantic unit in the set of clause-level semantic units with subject category labels is projected onto the corresponding coordinate position on the horizontal axis according to the message sending timestamp, and onto the corresponding coordinate layer on the vertical axis according to the subject category label. This ensures that each clause-level semantic unit occupies a unique coordinate position in the two-dimensional discourse flow space. Figure 3 Circular nodes on each coordinate layer. The layout, placing the ontological feature flow coordinate layer in the upper region of the vertical axis and the environmental noise flow coordinate layer in the lower region of the vertical axis, ensures that, within the visual structure of the two-dimensional discourse flow space, information with high to low relevance to the patient's condition is arranged from top to bottom. Figure 3 The three coordinate layers are arranged in the same order. The processing focus in step S20 is concentrated on the upper ontology feature flow coordinate layer and the middle object interaction flow coordinate layer. The clause-level semantic units in the lower environmental noise flow coordinate layer do not participate in the subsequent medical entity scanning and entity aggregation processing.

[0023] After projecting the clause-level semantic units, cross-level dependency pointers are established for clause-level semantic units originating from the same message slice but belonging to different coordinate layers, for association purposes. Figure 3Circular nodes at the same horizontal axis position in different coordinate layers. A cross-layer dependency pointer is a directed link from a clause-level semantic unit in one coordinate layer to a clause-level semantic unit in another coordinate layer. The condition for establishing the link is that the two clause-level semantic units carry the same timestamp label and originate from the same message slice. The physical meaning of the cross-layer dependency pointer is that there is an original contextual correlation between clause-level semantic units assigned to different coordinate layers after the clause-level semantic decomposition process of the same message slice in step S11. Although the logical subjects of the clause-level semantic units are different, the fact that clause-level semantic units co-occur in the same message indicates that the patient may have been influenced by the context of other people's information in the same message when expressing their own information, or there is a topical continuity between their own information and other people's information. In step S23, the cross-layer dependency pointer is used to detect clause-level semantic units belonging to the ontology feature flow coordinate layer within the time range covered by the shielding protection buffer distance interval on the horizontal axis. This allows step S23 to trace back from the object interaction flow coordinate layer to clause-level semantic units in the same message slice along the cross-layer dependency pointer, and to perform subject-verb-object relationship re-verification on medical entities within the clause-level semantic units. This enables a more rigorous confirmation of the attribution of medical entities in the patient's self-report information within the time interval where other people's information appears densely. Without the cross-layer dependency pointer, step S23 will not be able to locate clause-level semantic units in the ontology feature flow coordinate layer that are homologous to clause-level semantic units in the object interaction flow coordinate layer within the shielding protection buffer distance interval. The subject-verb-object relationship re-verification will not be able to focus specifically on clause-level semantic units in the ontology feature flow that have contextual co-occurrence relationships with other people's information. This results in either a verification range that is too narrow and misses clause-level semantic units that need to be verified, or a range that is too wide and increases unnecessary verification overhead.

[0024] Step S13 merges the set of clause-level semantic units with timestamps and ordinal numbers generated in step S11 with the subject category labels generated in step S12 into a two-dimensional spatial structure with time and attribution dimensions. This allows the originally linearly arranged clause-level semantic units to achieve a spatially hierarchical expression. The establishment of the two-dimensional discourse flow space enables step S20 to execute different processing strategies on different coordinate layers: calculating the medical entity density parameter for the object interaction flow coordinate layer to identify the time window to be shielded, and performing subject-verb-object relationship re-verification for the ontology feature flow coordinate layer within the shielding protection buffer distance interval of the time window to be shielded. Without step S13, the subject category labels generated in step S12 would only exist as attribute markers for clause-level semantic units and would not be able to form a spatial distribution structure in the time dimension. The operation of dividing the object interaction flow coordinate layer into continuous equal-width time windows according to the horizontal time dimension in step S20 would lack a spatial carrier, and the calculation of the medical entity density parameter would not be possible at the granularity of the time window.

[0025] Step S14: Perform structured parsing on the patient's discharge summary document, extract the diagnostic name entity, biochemical indicator entity, medication plan entity, and treatment operation entity to form a structured baseline dataset for the discharge summary. Perform medical entity pre-scanning on the clause-level semantic units in the ontology feature flow coordinate layer of the two-dimensional discourse flow space with cross-layer dependency pointers. Anchor the extracted medical entities to the structured baseline dataset of the discharge summary to obtain ontology feature flow coordinate layer data with baseline anchoring mapping.

[0026] See Figure 4 This is a schematic diagram of baseline anchoring mapping provided in the embodiments of this application. Figure 4 The diagram illustrates the correspondence between the structured baseline dataset of the discharge summary and the ontology feature flow coordinate layer, demonstrating the bidirectional mapping structure of four types of medical entities. The dashed arrows represent anchored mapping links. Specifically, in step S14, structured parsing processing is performed on the patient's discharge summary document. Structured parsing processing refers to extracting structured fields from the text content of the patient's discharge summary document according to predefined medical entity categories, forming... Figure 4 The discharged patient summary structured baseline dataset is shown. Predefined medical entity categories include... Figure 4The entities are categorized into four types: diagnostic name entities, biochemical indicator entities, medication regimen entities, and treatment operation entities, arranged sequentially. Structured parsing processing employs a combination of rule-based template matching and named entity recognition: For fields with fixed-format titles in the patient discharge summary document, rule-based template matching is used for extraction. This involves pre-defining common field title formats in discharge summary documents; when a text paragraph matching the predefined title format appears, the content following the title is extracted as the corresponding category of medical entity. For medical entities scattered throughout the text, named entity recognition is used for extraction. This involves using a conditional random field-based sequence labeling model trained on a medical corpus to perform word-by-word annotation, identifying named entities belonging to predefined medical entity categories and their boundaries. This conditional random field-based sequence labeling model uses the BIO annotation system, taking word segmentation sequences and part-of-speech tagging sequences as input features, and outputting B, I, or O tags for each word position. Named entities and their boundaries are obtained by extracting word segments covered by consecutive B and I tags. The extracted diagnostic name entities include the clinically confirmed disease name and diagnosis code of the patient at discharge; biochemical indicator entities include the indicator names and values ​​from laboratory test results such as blood tests and urine tests; medication regimen entities include the names, dosages, and methods of administration of medications to be taken home at discharge; and treatment operation entities include the names of surgeries, physical therapy plans, and nursing interventions performed during hospitalization. These four types of medical entities together constitute the structured baseline dataset for the discharge summary, which represents the clinically validated medical status baseline of the patient at the time of discharge.

[0027] After extracting the structured baseline dataset of discharge summaries, the two-dimensional discourse flow space with cross-layer dependency pointers was analyzed. Figure 4The document describes the pre-scanning of medical entities within the clause-level semantic units of the ontology feature flow coordinate layer. This pre-scanning involves scanning the text content of each clause-level semantic unit within the ontology feature flow coordinate layer using a combination of medical terminology dictionary matching and a conditional random field-based sequence labeling model. The goal is to identify medical entities such as symptom entities, drug name entities, treatment operation entities, and biochemical indicator entities that appear within these clause-level semantic units. The medical terminology dictionary matching process involves comparing the text content of sentence-level semantic units word-by-word with a pre-built medical terminology dictionary. When a word or phrase in the text of a sentence-level semantic unit matches an entry in the medical terminology dictionary, the matched word or phrase is identified as a candidate medical entity. The conditional random field-based sequence labeling model is a statistical sequence labeling model trained on medical corpora. This model uses the word segmentation sequence of sentence-level semantic units and the corresponding part-of-speech tagging sequence as input features. It uses a conditional random field to output a label for each word position in the input sequence. The label uses the BIO labeling system, where B labels represent the starting word of a medical entity, I labels represent internal continuation words of a medical entity, and O labels represent non-medical entity words. Medical entities and their boundaries are obtained by extracting word segments covered by consecutive B and I labels. The medical terminology dictionary matching process accurately identifies medical entities appearing in standard medical terminology, while the conditional random field-based sequence labeling model identifies medical entities expressed in colloquial or non-standard forms. The combination of these two methods improves the recall rate of pre-scanning medical entities. The medical entities extracted from the pre-scan medical entity model are anchored and mapped to the structured baseline dataset of the discharge summary: the semantic similarity of each medical entity extracted from the pre-scan medical entity model is compared with the four types of medical entities in the structured baseline dataset of the discharge summary. The semantic similarity is calculated by inputting the textual descriptions of the medical entities extracted from the pre-scan medical entity model and the textual descriptions of the medical entities in the structured baseline dataset of the discharge summary medical entity model into a semantic vector encoding model pre-trained with medical corpus to obtain their respective semantic vector representations, and then calculating the cosine similarity between the two semantic vector representations.The semantic vector encoding model pre-trained on the medical corpus uses a BERT pre-trained language model based on the Transformer encoder architecture as its base model. This BERT pre-trained language model contains multiple bidirectional Transformer encoder layers. Each Transformer encoder layer includes a multi-head self-attention sublayer and a feedforward neural network sublayer. The input text is segmented into a word sequence, which is then stacked with word embedding layers, position embedding layers, and segment embedding layers to obtain the input vector sequence. This input vector sequence is then sequentially processed by the multi-head self-attention calculation and feedforward neural network transformation of each Transformer encoder layer. The hidden state vector corresponding to the starting marker position in the final layer output is taken as the semantic vector representation of the input text. After completing the masked language model pre-training task on general Chinese corpora, this BERT pre-trained language model is further pre-trained on Chinese medical domain corpora including electronic medical records, clinical guidelines, drug instructions, and medical textbooks. The further pre-training task also adopts the masked language model task. By randomly masking input words in the medical domain corpus and training the model to predict the masked words, the semantic vector representation space of the model is adapted to the terminology distribution and semantic relationships in the medical domain, so that synonymous or near-synonymous medical terms have a high cosine similarity in the semantic vector space. The cosine similarity value is used as the semantic similarity between two medical entities. When the semantic similarity between a medical entity extracted by the pre-scan and a medical entity in the structured baseline dataset of the discharge summary is greater than or equal to the anchoring mapping similarity threshold, an anchoring mapping similarity is established between the two. Figure 4 The anchoring mapping links indicated by the dashed arrows mark the medical entities extracted from the pre-scan as anchored medical entities with baseline correspondence. Determining the anchoring mapping similarity threshold requires considering the degree of difference between the patient's spoken expression and standard medical terminology: if the anchoring mapping similarity threshold is set too high, medical entities described by the patient in a colloquial manner will fail to establish anchoring mapping links due to significant literal differences from the standard terminology in the structured baseline dataset of the discharge summary, resulting in the omission of a large number of valid anchored medical entities; if the anchoring mapping similarity threshold is set too low, medical entity pairs with weak semantic relevance will be incorrectly anchored, reducing the reliability of anchor matching in subsequent step S24. When determining the anchor mapping similarity threshold, common medical terms in the patient's discharge summary document are used as a reference. The semantic similarity distribution between colloquial expressions of the same medical concepts and standard terms in the original text of the patient's daily conversation records is statistically analyzed. The quantile value that can cover most (e.g., 85% to 95%) correct mapping pairs and exclude most incorrect mapping pairs in the semantic similarity distribution is selected as the anchor mapping similarity threshold.

[0028] After completing the anchoring mapping, the ontology feature flow coordinate layer data with baseline anchoring mapping is obtained. This data includes each clause-level semantic unit within the ontology feature flow coordinate layer, as well as the medical entities identified through pre-scanning of medical entities within those clause-level semantic units. Figure 4 The dashed arrows represent the anchoring mapping links between the medical entities and the structured baseline dataset of the discharge summary. Step S14 introduces the clinically validated medical facts from the patient's discharge summary document into the ontology feature flow coordinate layer of the two-dimensional discourse flow space, enabling the medical entities within the ontology feature flow coordinate layer to obtain baseline references from the structured baseline dataset of the discharge summary. In step S24, the ontology feature flow coordinate layer data with baseline anchoring mappings is used for anchor point matching with the confirmed set of ontology medical entities. The anchoring mapping links allow step S24 to compare the medical entities mentioned in the patient's daily conversations after discharge with the baseline of the medical state at the time of discharge, extracting a set of trend features before and after intervention. Without the baseline anchoring mappings in step S14, step S24 would be unable to distinguish between new symptoms after discharge and symptoms already present at discharge, nor would it be able to identify the continuity or change relationship between the discharge medication plan and the medication information mentioned in the patient's daily conversations. The set of trend features before and after intervention in the structured case summary text would lack the reference dimension of the discharge baseline.

[0029] Step S10 transforms the original text of the patient's daily conversation records from unstructured continuous text into a structured set with clause-level semantic units as the smallest granularity and timestamps and ordinal numbers, through the two-layer decomposition process in Step S11. This ensures that each clause-level semantic unit contains only semantic content pointed to by a single logical subject, eliminating the interference of multiple logical subjects mixed within the same text segment for subsequent subject category label assignment. Step S12's logical subject identification process, based on the clean input granules provided in Step S11, categorizes clause-level semantic units into three types according to logical subject attribution: ontology feature flow labels, object interaction flow labels, and environmental noise flow labels. This three-category system connects with the utilization of the medical entity density parameter in the object interaction flow coordinate layer in Step S20. The independent setting of the object interaction flow coordinate layer allows Step S20 to locate the time window where patient self-report information is prone to attribution confusion, without contaminating the ontology feature flow coordinate layer data, by utilizing the temporal distribution characteristics of other people's relevant medical information, thus achieving indirect localization of spoken language denoising. Step S13 combines the time dimension and the attribution dimension into a two-dimensional discourse flow space. The establishment of cross-layer dependency pointers maintains the original context tracing channel between different coordinate layers. Step S23 uses cross-layer dependency pointers to locate the semantic units at the sentence level of the ontology feature flow coordinate layer that need to be re-verified within the shielded protection buffer distance range. This mechanism transforms spoken language denoising from overall screening to targeted verification, retaining correctly attributed entities while eliminating incorrectly attributed entities, thus avoiding the loss of effective medical information due to over-screening. Step S14 anchors the clinically verified medical facts in the patient's discharge summary document to the ontology feature flow coordinate layer in the form of a structured baseline dataset of the discharge summary. This allows Step S24 to compare the patient's self-reported medical information after discharge with the baseline of medical status at the time of discharge when generating the structured case summary text, extracting a set of trend features before and after intervention. There is a granularity matching relationship between the two-layer decomposition process in step S11 and the logical subject identification process in step S12. Step S11 refines the input granularity to the semantic unit at the clause level, so that the logical subject identification in step S12 can operate on the text fragment pointed to by a single subject. If step S11 only performs the first-layer timestamp slice processing without performing the second-layer clause-level semantic decomposition processing, step S12 will have to perform logical subject identification at the message slice granularity containing multiple logical subjects. At this time, the logical subject identification result can only assign a subject category label to the entire message slice. Semantic fragments in the same message slice that point to different objects will be forced to be classified into the same coordinate layer. The calculation of the medical entity density parameter and the re-verb-object relationship in step S20 will produce a judgment bias due to the mixing of medical entities of different affiliations in the coordinate layer.In step S13, the cross-layer dependency pointer and in step S14, the baseline anchoring mapping, play complementary roles in step S20: the cross-layer dependency pointer provides a contextual tracing path across coordinate layers, enabling step S23 to find the semantic units at the sentence level of the ontology feature flow coordinate layer that need to be re-verified within the shielded protection buffer distance interval; the baseline anchoring mapping provides an entity-level reference for the discharge baseline, enabling step S24 to distinguish between known existing medical entities and newly emerging incremental medical entities in anchor point matching. The combination of the two ensures that the structured case summary text excludes medical entities with incorrect attribution while retaining complete information on the trends before and after intervention. Step S10 transforms the original texts of patients' daily dialogue records from different sources and with vastly different structures, along with the patient's discharge summary document, into a unified spatial structure with time and attribution dimensions. A cross-layer tracing channel and a baseline anchoring reference are established within this spatial structure, allowing subsequent spoken language denoising processing to operate at the granularity of targeted verification and ensuring that the generation of the structured case summary text is supported by clinically validated discharge baselines.

[0030] Step S20: Based on the two-dimensional discourse flow space with cross-layer dependency pointers, calculate the density parameter of medical entities in the object interaction flow coordinate layer of the two-dimensional discourse flow space, identify the time window to be shielded based on the density parameter of medical entities, and generate the shielding protection buffer distance interval for each time window to be shielded. Within the shielding protection buffer distance interval, perform subject-verb-object relationship re-verification processing on the medical entities in the ontology feature flow coordinate layer. After verifying the confirmed ontology medical entity set, aggregate the ontology feature flow coordinate layer data with baseline anchoring mapping and the structured baseline dataset of discharge summary, generate the structured case closure summary text.

[0031] Specifically, step S20, based on the two-dimensional discourse flow space with cross-layer dependency pointers constructed in step S10 and the ontology feature flow coordinate layer data with baseline anchoring mapping, performs directional identification and denoising processing on non-patient medical entities mixed in the spoken text. The denoised and confirmed patient-related medical entities are then aggregated with the structured baseline dataset of the discharge summary to generate a structured case closure summary text. When a large number of medical entities related to others appear in the original text of a patient's daily conversation within a certain time interval, the patient's self-described medical entities within the same time interval have a high risk of misattribution. That is, the patient may be relaying information about others while indirectly describing others' symptoms, which could be misidentified as the patient's own symptoms. Therefore, more stringent attribution verification is needed for these high-risk time intervals of patient-described medical entities. Step S20 quantifies the distribution characteristics of other people's relevant medical information in the time dimension by calculating the density parameter of medical entities on the object interaction flow coordinate layer, thereby identifying the time window in which other people's medical information is concentrated and needs to be shielded. Within the shielding time window and the surrounding shielding protection buffer distance interval, the subject-verb-object relationship of the medical entities in the ontology feature flow coordinate layer is re-verified. Finally, the confirmed ontology medical entity set after verification is aggregated with the structured baseline dataset of discharge summary, and the structured case closure summary text is output.

[0032] Further, step S20 includes: Step S21: Perform medical entity scanning processing on all clause-level semantic units in the object interaction flow coordinate layer of the two-dimensional discourse flow space with cross-layer dependency pointers, identify symptom entities, drug name entities, treatment operation entities, and biochemical indicator entities, and obtain the object interaction flow medical entity list. Divide the object interaction flow coordinate layer into continuous equal-width time windows according to the horizontal axis time dimension, calculate the ratio of the number of medical entities recorded in the object interaction flow medical entity list to the total number of clause-level semantic units in each equal-width time window, and obtain the medical entity density parameter for each time window.

[0033] See Figure 5 This is a schematic diagram of the density of medical entities within an equal-width time window provided in the embodiments of this application. Figure 5 The diagram illustrates the equal-width time window division structure along the time axis of the object interaction flow coordinate layer, demonstrating the differences in the distribution of medical entities within different time windows. Each circular node represents a medical entity. Specifically, in step S21, the two-dimensional discourse flow space with cross-layer dependency pointers... Figure 5All clause-level semantic units within the object interaction flow coordinate layer are subjected to medical entity scanning processing. This medical entity scanning process employs the same technical approach as the medical entity pre-scanning in step S14, combining medical terminology dictionary matching with a conditional random field-based sequence labeling model. It scans the text content of each clause-level semantic unit within the object interaction flow coordinate layer one by one, identifying medical entities such as symptom entities, drug name entities, treatment operation entities, and biochemical indicator entities. Symptom entities refer to medical terms describing physical discomfort or abnormal sensations; drug name entities refer to the generic or brand name of a drug; treatment operation entities refer to the names of examinations, surgeries, or interventions performed during clinical treatment; and biochemical indicator entities refer to the names of quantifiable parameters in laboratory test results such as blood tests and urine tests. All identified medical entities are summarized according to their affiliation with clause-level semantic units to form an object interaction flow medical entity list. This list records the timestamp and sequence number information of each identified medical entity within the object interaction flow coordinate layer and the clause-level semantic unit to which it belongs. The reason for performing medical entity scanning processing on the object interaction flow coordinate layer rather than the ontology feature flow coordinate layer is that the object interaction flow coordinate layer carries medical information related to others. The denoising logic in step S20 uses the density of medical information related to others in the time dimension to locate the time interval in the patient's self-report information where there may be confusion about attribution. Performing medical entity scanning processing on the object interaction flow coordinate layer can obtain the spatial distribution characteristics of medical information of others. This distribution characteristic is the data source for subsequent calculation of the medical entity density parameter.

[0034] After completing the construction of the medical entity list for the object interaction flow, the object interaction flow coordinate layer is divided into sections according to the horizontal time dimension. Figure 5 The diagram shows continuous, equal-width time windows. An equal-width time window refers to a horizontal axis in a two-dimensional discourse flow space, where the entire time range covered by the axis is evenly divided into several consecutive, non-overlapping time segments with a fixed time interval. Figure 5The example demonstrates four consecutive equal-width time windows. The width of the equal-width time window needs to cover the topic transition cycles that a patient may be involved in during a continuous conversation, ensuring that each equal-width time window contains a sufficient number of clause-level semantic units to reflect the density of others' medical information within a local time period. Determining the width of the equal-width time window requires considering two factors: the message sending frequency of the patient's daily conversations and the duration of the topics. If the width of the equal-width time window is too narrow, the number of clause-level semantic units contained within a single equal-width time window will be too small, and the calculation result of the medical entity density parameter may easily produce false high-density values ​​due to medical entities appearing occasionally in a single message, leading to an increased false positive rate triggered by blocking. If the width of the equal-width time window is too wide, it will span multiple topic transition cycles, diluting the local time periods where others' medical information appears intensively by a large number of non-intensive time periods, causing the medical entity density parameter to fail to reflect local density peaks, thus increasing the false negative rate of the blocked time window. When determining the width of the equal-width time window, the average interval between message timestamps in the original text of the patient's daily conversation is used as a benchmark. The width of the equal-width time window is selected as the time span covering several consecutive messages. This ensures that the number of clause-level semantic units included in the window maintains responsiveness to local topic shifts while suppressing occasional fluctuations in individual messages. For example, when the average interval between message timestamps in the original text of the patient's daily conversation is several minutes, the width of the equal-width time window can be set to cover a time span of 10 to 20 messages.

[0035] After dividing the time windows into equal-width segments, the density parameter of medical entities for each equal-width time window is calculated. The calculation method for the density parameter is as follows: Count the number of medical entities recorded in the medical entity list of the object interaction flow within the time range covered by each equal-width time window on the horizontal axis. Simultaneously, count the total number of clause-level semantic units contained in the object interaction flow coordinate layer within the time range covered by the same equal-width time window on the horizontal axis. Divide the number of medical entities by the total number of clause-level semantic units; the resulting ratio is the density parameter of medical entities for the equal-width time window. Figure 5The second time window contains multiple medical entities, and its medical entity density parameter is significantly higher than other time windows, indicating a high density of medical entities. When the object interaction flow coordinate layer within a certain equal-width time window does not contain any clause-level semantic units (i.e., the total number of clause-level semantic units is zero), the medical entity density parameter for that equal-width time window is set to zero, indicating that there is no medical information related to others within that time window, and the blocking process is not triggered. The lower limit of the medical entity density parameter is 0, and theoretically there is no strict upper limit constraint. When none of the clause-level semantic units in the object interaction flow coordinate layer within the equal-width time window contain medical entities, the medical entity density parameter is 0; when each clause-level semantic unit contains at least one medical entity, the medical entity density parameter is equal to or greater than 1. However, in actual spoken text, since the text content of clause-level semantic units is mainly daily dialogue, most clause-level semantic units are short and do not contain medical entities, so the medical entity density parameter is usually much less than 1, and cases greater than or equal to 1 are extremely rare. The reason for using the ratio of the number of medical entities to the total number of clause-level semantic units, rather than the absolute value of the number of medical entities, to measure density is that the number of clause-level semantic units contained in different equal-width time windows may vary due to differences in the frequency of messages sent by patients. If only the absolute value of the number of medical entities is used as the indicator of density, equal-width time windows with higher message sending frequencies will have a higher absolute number of medical entities due to the large number of clause-level semantic units. Even if the proportion of medical entities appearing in clause-level semantic units is not high, it may be misjudged as a dense interval. Using the ratio eliminates the interference of differences in message sending frequency on the determination of density. Without step S21, which involves scanning and processing medical entities at the object interaction flow coordinate layer and calculating the density parameters of medical entities, step S22 will not be able to obtain quantitative data on the distribution of others' medical information in the time dimension, and the identification of the masked time window will lack data support.

[0036] Step S22: Set a shielding trigger threshold parameter and compare the density parameter of medical entities in each time window with the shielding trigger threshold parameter to determine and filter out the time windows that need to be shielded. Apply a shielding mask mark to the medical entity list of the object interaction flow for the time windows that need to be shielded, and extend along the horizontal axis with the start and end timestamps of the time windows that need to be shielded as the center boundary to generate the shielding protection buffer distance range for each time window that needs to be shielded.

[0037] See Figure 6 This is a schematic diagram of the shielding protection buffer distance range provided in the embodiments of this application. Figure 6The diagram illustrates the shielding and protection buffer distance interval structure distributed along the time axis, showing the hierarchical division of the core area of ​​the time window to be shielded and the preceding and following extended buffer zones, and marking the boundary positions of the start and end timestamps. Specifically, in step S22, a shielding trigger threshold parameter is established, and the density parameter of medical entities in each equal-width time window obtained in step S21 is numerically compared with the shielding trigger threshold parameter. The shielding trigger threshold parameter is a value between 0 and 1, used to define the level at which the density of other people's related medical information in the object interaction flow coordinate layer needs to reach to trigger the attribution verification of the patient's self-reported medical entities in adjacent time intervals in the ontology feature flow coordinate layer. When the density parameter of medical entities in a certain equal-width time window is greater than or equal to the shielding trigger threshold parameter, this equal-width time window is determined to be... Figure 6 The solid-line box marks the time windows that need to be masked. When the density parameter of medical entities in a certain equal-width time window is less than the masking trigger threshold parameter, this equal-width time window will not trigger masking processing. The determination of the masking trigger threshold parameter needs to consider two factors: the actual probability of confusion between the attribution of others' medical information and the patient's self-reported information in spoken text, and the computational cost of verification processing. If the masking trigger threshold parameter is set too low, a large number of equal-width time windows with low density of others' medical information will be judged as time windows that need to be masked, triggering a re-verification of the subject-verb-object relationship of a large range of clause-level semantic units in the ontology feature flow coordinate layer, increasing unnecessary computational overhead. Moreover, the probability of confusion between the attribution of the patient's self-reported information is low in time intervals with low density of others' medical information, so the benefit of performing verification on these intervals is limited. If the masking trigger threshold parameter is set too high, only equal-width time windows with extremely dense others' medical information will trigger masking processing. A large number of time intervals with moderately dense others' medical information will be missed. There is still a risk of attribution confusion in the patient's self-reported information in these moderately dense intervals. Missing these intervals will result in some incorrectly attributed medical entities not being verified and entering the confirmed ontology medical entity set. When determining the masking trigger threshold parameter, a labeled analysis is performed on historical samples of the original text of patients' daily conversations. The frequency distribution of patient-reported medical entity attribution errors is statistically analyzed under different levels of medical entity density parameters. The medical entity density parameter value corresponding to the inflection point where the frequency of attribution errors changes from low to high is selected as the masking trigger threshold parameter, ensuring that the masking process covers the density range where the risk of attribution errors jumps from low to high. For example, the masking trigger threshold parameter can be set to a value between 0.3 and 0.5.

[0038] After determining the time windows requiring masking, a masking marker is applied to the object interaction flow medical entity list for each equal-width time window deemed to require masking. The masking marker is a status label used to identify entities in the object interaction flow medical entity list that are located in... Figure 6Medical entities in the core area have been included in the masking process. These medical entities carrying masking markers will not participate in the merging process with medical entities in the ontology feature flow coordinate layer during the entity aggregation process in subsequent step S24, thus preventing other medical entities from being incorrectly introduced into the structured case summary text. After applying the masking markers, the start and end timestamps of each masking time window on the horizontal axis are used as the center boundary, extending forward and backward along the horizontal axis for a certain period of time, forming... Figure 6 The shielding protection buffer distance interval is the area jointly covered by the dashed and solid bounding boxes for each time window requiring shielding. The time coverage range of the shielding protection buffer distance interval is greater than the time coverage range of the time window requiring shielding itself. Figure 6 The buffer extends forward to cover the time period before the time window to be blocked, and extends backward to cover the time period after the time window to be blocked. The reason for setting the buffer distance range is that topic transitions in patients' daily conversations are not instantaneous. When a patient intensively discusses others' medical information within a certain time period, their statements may still be influenced by the topics of others' medical information in the adjacent time periods before and after this period. While the risk of attribution confusion for the medical entities described by the patient is lower in these adjacent time periods than within the time window to be blocked, it is higher in periods far from the time window to be blocked. The extension length of the buffer distance range needs to cover the transitional area of ​​topic switching: if the extension length is too short, clause-level semantic units near the boundary of the time window to be blocked, still influenced by others' medical information, will fall outside the buffer distance range and not be verified; if the extension length is too long, a large number of clause-level semantic units far from the time window to be blocked and with extremely low attribution confusion risk will be included in the verification scope, increasing unnecessary computational overhead. When determining the extension length of the shielding protection buffer distance interval, the width of the equal-width time window is used as the reference unit, and a certain multiple (such as 1 to 2 times) of the width of the equal-width time window is selected as the extension length, so that the buffer interval covers the range of one to two equal-width time windows before and after the time window to be shielded. If step S22 is missing the determination of the time window to be shielded and the generation of the shielding protection buffer distance interval, step S23 will not be able to determine the time range for subject-verb-object relationship re-verification. Subject-verb-object relationship re-verification will either be performed indiscriminately on all sentence-level semantic units of the ontology feature flow coordinate layer, resulting in a large amount of unnecessary computational overhead, or it will not be performed at all, making it impossible to identify and exclude medical entities with incorrect attribution. Step S22 transforms the continuously distributed medical entity density parameters obtained in step S21 into a discrete set of time windows to be shielded and a corresponding set of shielding protection buffer distance intervals, so that the subject-verb-object relationship re-verification in step S23 can focus on the time interval with the highest risk of attribution confusion, achieving a balance between verification coverage and computational cost.

[0039] Step S23: Based on the cross-layer dependency pointer detection shielding protection buffer distance interval, within the time range covered by the horizontal axis, the clause-level semantic unit belonging to the ontology feature flow coordinate layer is subjected to subject-verb-object relation re-verification of the medical entity in the clause-level semantic unit, the attribution status of the medical entity is determined, and the ontology feature flow coordinate layer data after subject-verb-object relation re-verification is obtained.

[0040] Specifically, see Figure 7In step S23, based on the cross-layer dependency pointer established in step S13, the sentence-level semantic units belonging to the ontology feature flow coordinate layer within the time range covered by each shielding protection buffer distance interval on the horizontal axis are detected. The detection method is as follows: traverse all sentence-level semantic units of the object interaction flow coordinate layer within the time range covered by the shielding protection buffer distance interval on the horizontal axis, trace back to the ontology feature flow coordinate layer along the cross-layer dependency pointer attached to each sentence-level semantic unit, and obtain the ontology feature flow coordinate layer sentence-level semantic units that share the same timestamp label and originate from the same message slice as the sentence-level semantic units of the object interaction flow coordinate layer. These ontology feature flow coordinate layer sentence-level semantic units obtained through tracing via cross-layer dependency pointers are patient self-narrative sentence-level semantic units that have original message co-occurrence relationships with other people's information in the context of dense occurrence of other people's medical information. The medical entities in these sentence-level semantic units have a high risk of attribution confusion. Subject-verb-object relationship re-verification is performed on the medical entities in each ontology feature flow coordinate layer sentence-level semantic unit obtained through tracing. Subject-verb-object relation re-verification refers to performing word segmentation and part-of-speech tagging on the text of clause-level semantic units to obtain a part-of-speech tagging sequence, then performing dependency parsing on the part-of-speech tagging sequence to obtain a dependency syntax tree, extracting the grammatical relationship path between the medical entity and the predicate verb in the dependency syntax tree, determining the grammatical role of the medical entity in the statement, and the semantic association between the medical entity and the logical subject. When there is no direct subject-verb-object dependency path between the medical entity and the main predicate verb of the clause-level semantic unit in the dependency syntax tree, or when a dependency path exists between the medical entity and the main predicate verb of the clause-level semantic unit but this path passes through a subordinate predicate verb in a nested clause or indirect quotation structure and the logical subject of the nested clause points to another, the attribution state of the medical entity is determined to be invalid. This indicates that although the medical entity appears in the clause-level semantic unit at the ontology feature flow coordinate layer, it is not directly related to the main logical subject of the clause-level semantic unit in the grammatical structure of the dependency syntax tree. In a dependency syntax tree, a medical entity is considered to belong to the patient if it is directly subject-verb-object dependent on the main verb of a clause-level semantic unit. This dependency path does not pass through any subordinate verb of a nested clause or indirect quotation structure, or if it passes through a subordinate verb of a nested clause but the logical subject of the nested clause points to the patient. In this case, the medical entity is considered to belong to the patient.When there is no direct subject-verb-object dependency path between a medical entity and its logical subject in the dependency syntax tree, or when the dependency path of a medical entity passes through a subordinate verb in a nested clause or indirect quotation structure and the logical subject of the nested clause points to another person, the attribution status of the medical entity is determined to be invalid. This means that although the medical entity appears in the clause-level semantic unit of the ontology feature flow coordinate layer, it should not actually belong to the patient. After completing the subject-verb-object relation re-verification, the ontology feature flow coordinate layer data after subject-verb-object relation re-verification is obtained. Each medical entity in the ontology feature flow coordinate layer data after subject-verb-object relation re-verification carries a valid attribution status marker or an invalid attribution status marker.

[0041] Step S23 locates the semantic units at the clause level of the ontology feature flow coordinate layer that need to be verified through cross-layer dependency pointers. This ensures that the subject-verb-object relationship re-verification is performed only on those semantic units at the clause level of the ontology feature flow coordinate layer that have a message co-occurrence relationship with the semantic units at the object interaction flow coordinate layer, rather than indiscriminately performing the verification on all semantic units at the clause level of the ontology feature flow coordinate layer. This targeted verification mechanism narrows the verification scope from all semantic units at the clause level of the ontology feature flow coordinate layer to a subset of semantic units at the clause level with cross-layer dependency pointer associations within the shielded protection buffer distance interval. While reducing computational overhead, it concentrates verification resources on the semantic units at the clause level with the highest risk of attribution confusion. For semantic units at the clause level of the ontology feature flow coordinate layer outside the shielded protection buffer distance interval, and for semantic units at the clause level of the ontology feature flow coordinate layer within the shielded protection buffer distance interval but without cross-layer dependency pointer associations, step S23 retains the original judgment result of the logical subject identification processing in step S12, does not perform subject-verb-object relationship re-verification, and defaults the attribution status of the medical entity to a valid attribution status. The reason for not performing subject-verb-object relation re-verification on this type of clause-level semantic unit is that the absence of cross-layer dependency pointer association means that there are no clause-level semantic units assigned to the object interaction flow coordinate layer in the message slice to which the clause-level semantic unit belongs. That is, all clause-level semantic units in the message slice are determined by step S12 to be patient self-report information. There is no contextual condition for co-occurrence of patient self-report information and other people's information within the same message slice. Therefore, the risk of medical entities in this clause-level semantic unit being misattributed due to the mixing of subjects in the message is significantly lower than that of clause-level semantic units with cross-layer dependency pointer association. The benefit of performing subject-verb-object relation re-verification on it is not enough to offset the increased computational overhead. If the subject-verb-object relationship re-verification in step S23 is missing, the classification result based on the overall logical subject of the clause-level semantic unit in step S12 will be directly used in step S24. For other medical entities that appear in the ontology feature flow coordinate layer but are actually introduced by nested clauses or indirect quotations, the logical subject identification at the clause-level semantic unit granularity in step S12 cannot detect the finer-grained subject switching within the clause-level semantic unit. These medical entities will enter the confirmed ontology medical entity set with a valid attribution status and will eventually be written into the structured case summary text, causing the attribution of patient medical information to be incorrect. The subject-verb-object relationship re-verification in step S23 adds a medical entity-level attribution determination to the logical subject identification at the clause-level semantic unit granularity in step S12. The determination at the two granular levels forms a coarse-to-fine two-layer filtering structure. Step S12 diverts most of the other people's information to the object interaction flow coordinate layer at the clause-level semantic unit level. Step S23 performs a step-by-step identification of the few ambiguous entities remaining in the ontology feature flow coordinate layer at the medical entity level. The combination of the two layers of filtering makes the attribution reliability of medical entities entering the confirmed ontology medical entity set higher than the level that any single layer of filtering can achieve.

[0042] Step S24: See below for further details. Figure 7 All medical entities with valid attribution status labels are selected from the ontology feature flow coordinate layer data after subject-verb-object relation re-verification processing to obtain a confirmed ontology medical entity set. The confirmed ontology medical entity set is sorted according to the horizontal axis time dimension to form a temporal ontology medical entity sequence. The temporal ontology medical entity sequence is anchored and matched with the discharge summary structured baseline dataset and the ontology feature flow coordinate layer data with baseline anchoring mapping to extract the set of change trend features before and after intervention. The confirmed ontology medical entity set and the set of change trend features before and after intervention are processed to generate a structured case closure summary text.

[0043] Specifically, in step S24, all medical entities with valid attribution status markers are selected from the ontology feature flow coordinate layer data after subject-verb-object relation re-verification processing, resulting in a confirmed ontology medical entity set. Each medical entity in the confirmed ontology medical entity set has undergone two levels of attribution confirmation: logical subject identification processing in step S12 and subject-verb-object relation re-verification in step S23. This represents a medical entity extracted from the original text of the patient's daily dialogue records and, after speech denoising processing, confirmed to belong to the patient. All medical entities in the confirmed ontology medical entity set are sorted according to the horizontal time dimension, that is, arranged from earliest to latest according to the timestamp label of the clause-level semantic unit to which each medical entity belongs, forming a temporally sequenced ontology medical entity sequence. The temporally sequenced ontology medical entity sequence preserves the order of appearance of the patient's self-reported medical information in the time dimension, enabling subsequent trend analysis to identify the state evolution of the same type of medical entity at different time nodes based on the chronological relationship on the time axis.

[0044] Anchor point matching is performed between the temporally sequenced ontology medical entity sequence, the structured baseline dataset of discharge summaries, and the ontology feature flow coordinate layer data with baseline anchoring mappings. The anchor point matching process is as follows: for each medical entity in the temporally sequenced ontology medical entity sequence, the anchoring mapping link corresponding to the medical entity in the temporally sequenced ontology feature flow coordinate layer data with baseline anchoring mappings is found. This link is then used to trace back to the corresponding medical entity in the structured baseline dataset of discharge summaries to obtain the baseline state value at the time of discharge. For medical entities that appear in the temporally sequenced ontology medical entity sequence and also have corresponding anchoring mappings in the structured baseline dataset of discharge summaries, the baseline state value at the time of discharge is compared with the current state description mentioned in the patient's daily conversations. The direction and magnitude of change information are extracted to form a trend record in the pre- and post-intervention trend feature set. The extraction method for information on the direction and magnitude of change is as follows: For medical entities with explicit numerical expressions, such as biochemical indicators, when patients mention specific numerical information in daily conversations, the numerical value is directly compared with the baseline value of the corresponding indicator in the structured baseline dataset of the discharge summary. The direction of change (increase or decrease) and the absolute amount of the numerical difference are obtained as the magnitude of change. For medical entities without explicit numerical expressions, such as symptom entities and medication regimen entities, the direction and magnitude of change are determined by identifying degree modifiers and trend descriptive words that accompany the medical entity in the patient's daily conversations. Severity modifiers are words that describe the severity of symptoms or changes in medication dosage (e.g., "worsened," "much better," "reduced by half"). Trend descriptors are words that describe the direction of change in the state over time (e.g., "increasingly," "no longer," "starting again"). When severity modifiers and trend descriptors indicate worsening symptoms or new medication, the direction of change is recorded as deterioration. When severity modifiers and trend descriptors indicate symptom relief or reduced medication, the direction of change is recorded as improvement. When no obvious severity modifiers or trend descriptors appear in the patient's description, the direction of change is recorded as stable. The magnitude of change is categorized into three levels based on the intensity of the degree modifier: slight change, moderate change, and significant change. The specific grading rules are as follows: when the degree modifier expresses a weak or uncertain degree of change, the magnitude of change is judged as slight change, including but not limited to expressions such as "a little," "slightly," "seems," or "seems to be a bit better." When the degree modifier expresses a clear degree of change with a certain intensity, the magnitude of change is judged as moderate change, including but not limited to expressions such as "obvious," "much better," "worsened," or "halved." When the degree modifier expresses a strong degree of change or points to an extreme state, the magnitude of change is judged as significant change, including but not limited to expressions such as "completely," "thoroughly," "severely worsened," or "doesn't hurt at all."The set of trend features before and after intervention includes the state change trajectory of all medical entities with baseline correspondence from the time of discharge to the time of daily dialogue records. For medical entities that appear in the temporal ontology medical entity sequence but do not have a corresponding anchor mapping in the structured baseline dataset of the discharge summary, they are marked as newly added medical entities after discharge, indicating that the patient has new symptoms or new medication information not recorded at the time of discharge. These newly added medical entities after discharge are also included in the set of trend features before and after intervention and are marked as new items without baseline control.

[0045] After completing anchor point matching and extracting the feature set of changes before and after intervention, text generation processing is performed on the confirmed ontology medical entity set and the feature set of changes before and after intervention to obtain a structured case summary text. The text generation process is as follows: medical entities in the confirmed ontology medical entity set are grouped according to medical entity category, and within each group, they are arranged in chronological order according to the ontology medical entity sequence. Combined with the corresponding trend records in the feature set of changes before and after intervention, text paragraphs are generated in the structured format of a medical document. The structured case summary text includes all confirmed medical entities described by the patient after discharge, comparative change information between each medical entity and the discharge baseline, and information on newly added medical entities after discharge. Step S24 cross-references the attribution-confirmed medical entities output in step S23 with the baseline anchoring mapping established in step S14, so that each piece of medical information in the structured case summary text not only has the attribution reliability confirmed by double-layer filtering, but also has the change trend comparison information with the discharge baseline. Therefore, the structured case summary text can present the complete picture of the patient's medical status evolution from discharge to the present moment. Without step S24, the ontology feature flow coordinate layer data output from step S23, after subject-verb-object relation re-verification processing, would only contain medical entities and attribution status markers scattered across various time coordinates of the ontology feature flow coordinate layer in the two-dimensional discourse flow space. This information lacks temporal arrangement, baseline comparison, and classification summarization, making it impossible to form a structured document for clinical reference. Step S24 consolidates all intermediate processing results from steps S10 to S23 into a single structured text output, thus condensing the processing results of the dual-track entity recognition spoken language denoising parsing method for multi-source heterogeneous long texts into a deliverable document.

[0046] Step S20, through step S21, performs medical entity scanning processing on the object interaction flow coordinate layer and calculates the density parameter of medical entities within an equal-width time window. This quantifies the distribution characteristics of others' relevant medical information in the time dimension into a value comparable to a threshold, enabling step S22 to determine the time window to be blocked in a data-driven manner, rather than relying on manual experience or fixed rules to define high-risk time intervals. Based on the determined time window to be blocked, step S22 generates a blocking protection buffer distance interval, extending the blocking range forward and backward from the strict boundary of the time window to cover the transition interval of topic shifts. This allows the subject-verb-object relationship re-verification in step S23 to be performed within the complete gradient range of attribution confusion risk from high to medium, avoiding the omission of ambiguous entities near the boundary due to focusing only on the area within the time window to be blocked. Step S23 utilizes cross-layer dependency pointers to directionally trace the semantic units at the clause level of the ontology feature flow coordinate layer within the shielded protection buffer distance range. On top of the logical subject identification at the clause level semantic unit granularity in step S12, the subject-verb-object relationship attribution determination at the medical entity granularity is superimposed, forming a coarse-to-fine two-layer attribution filtering structure. This two-layer filtering structure allows clause level semantic units that have been initially screened and assigned to the ontology feature flow coordinate layer in step S12 but still contain other medical entities introduced by nested clauses or indirect quotations to be further identified in step S23. Medical entities with incorrect attribution are marked as invalid attribution status and excluded from the confirmed ontology medical entity set. The equal-width time window division in step S21 and the generation of the shielding protection buffer distance interval in step S22 form a transformation chain from continuous density distribution to discrete high-risk interval positioning. The shielding protection buffer distance interval in step S22 provides spatial range constraints for the cross-layer dependency pointer tracing in step S23. The three steps are connected in sequence, so that the execution range of the subject-verb-object relationship re-verification is neither too narrow to cause omissions nor too wide to cause redundancy, thus achieving a balance between verification coverage and computational cost. Step S24 performs anchor point matching between the confirmed ontology medical entity set output in step S23 and the structured baseline dataset of the discharge summary and the ontology feature flow coordinate layer data with baseline anchoring mapping provided in step S14. It extracts the set of change trend features before and after intervention and compares the patient's self-reported medical entities after verbal denoising with the clinical baseline at the time of discharge. This ensures that the structured case summary text not only contains the patient's current medical entity information but also the change trajectory from the discharge baseline to the current state. The generation of this change trajectory information depends on the anchoring mapping link established in step S14 of step S10. If step S14 does not establish an anchoring mapping between the medical entities in the ontology feature flow coordinate layer and the structured baseline dataset of the discharge summary in advance, step S24 will not be able to perform anchor point matching, and the structured case summary text will only contain the patient's current medical entity list and lack the change trend dimension.Step S20 identifies the time interval with the highest risk of attribution confusion in the patient's self-reported information by using the temporal distribution characteristics of other people's medical information in the object interaction flow coordinate layer. Attribution re-verification is performed at the medical entity granularity to exclude other people's medical entities remaining in the ontology feature flow coordinate layer. Finally, the patient's own medical entity after the double-layer attribution filtering is aggregated and compared with the discharge clinical baseline to generate a structured case summary text containing the change trend before and after the intervention.

[0047] Example 2: This embodiment, based on Embodiment 1, provides a dual-track entity recognition and spoken language denoising parsing system for multi-source heterogeneous long texts, such as... Figure 8 As shown, it includes: The decomposition space construction module is used to receive the original text of the patient's daily dialogue record and the patient's discharge summary document, perform two-layer decomposition on the original text of the patient's daily dialogue record, generate a set of clause-level semantic units with timestamp tags and sequence numbers, and construct a two-dimensional discourse flow space. The two-dimensional discourse flow space is divided into an ontology feature flow coordinate layer, an object interaction flow coordinate layer and an environmental noise flow coordinate layer on the vertical axis. Cross-layer pointer construction module: Based on the set of clause-level semantic units with timestamp tags and ordinal numbers and the two-dimensional discourse flow space, a two-dimensional discourse flow space with cross-layer dependency pointers is obtained; The shielding interval generation module: Based on the two-dimensional discourse flow space with cross-layer dependency pointers, it calculates the density parameters of medical entities in the object interaction flow coordinate layer of the two-dimensional discourse flow space, identifies the time windows that need to be shielded based on the density parameters of medical entities, and generates the shielding protection buffer distance interval for each time window that needs to be shielded. Entity Verification and Closure Module: This module is used to perform subject-verb-object re-verification processing on medical entities within the ontology feature flow coordinate layer within the shielded protection buffer distance range, and generate a structured closure summary text.

[0048] Furthermore, in the decomposition space construction module, the two-layer decomposition includes a first-layer timestamp slicing process and a second-layer clause-level semantic decomposition process; The first-level timestamp slicing process involves slicing the original text of the patient's daily conversation records according to the message sending timestamp, resulting in a message slice sequence arranged in chronological order. The second-level semantic decomposition process involves: for each message slice in the chronologically ordered message slice sequence, a joint delimitation mechanism using punctuation marks and conjunctions is used to divide it into several independent semantic units at the clause level; the timestamp label and sequence number are retained to construct a set of semantic units at the clause level with timestamp label and sequence number.

[0049] The method for constructing a two-dimensional discourse flow space includes: Based on the set of clause-level semantic units with timestamp labels and ordinal numbers, a set of clause-level semantic units with subject category labels is obtained; The timestamps of message slice sequences arranged in chronological order are extracted as the horizontal axis coordinates, and the subject category labels of the sentence-level semantic unit sets with subject category labels are extracted as the vertical axis coordinates to establish a two-dimensional discourse flow space.

[0050] The method for generating the set of clause-level semantic units with subject category labels includes: For each clause-level semantic unit in the set of clause-level semantic units with timestamp labels and ordinal numbers, logical subject identification processing is performed. Based on the logical subject identification results, each clause-level semantic unit is assigned one of three subject category labels: ontology feature flow label, object interaction flow label, or environmental noise flow label, resulting in a set of clause-level semantic units with subject category labels.

[0051] The method for generating the two-dimensional discourse flow space with cross-layer dependency pointers includes: Each clause-level semantic unit in the set of clause-level semantic units with subject category labels is projected to the corresponding coordinate position on the corresponding coordinate layer in the two-dimensional discourse flow space according to the timestamp and subject category label. Cross-layer dependency pointers are established for clause-level semantic units that originate from the same message slice but belong to different coordinate layers, and the two-dimensional discourse flow space with cross-layer dependency pointers is output.

[0052] The methods and systems of this application may be implemented in many ways. For example, they may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above-described order of steps for the method is for illustrative purposes only, and the steps of the method of this application are not limited to the order specifically described above, unless otherwise specifically stated.

[0053] In addition, the parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of the corresponding technical solutions in the prior art have not been described in detail, so as to avoid excessive elaboration.

[0054] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. 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 dual-track entity recognition spoken language denoising and parsing method for multi-source heterogeneous long texts, characterized in that, The method includes: The system receives the original text of patients' daily conversation records, performs a two-layer decomposition on the original text, generates a set of clause-level semantic units with timestamp tags and ordinal numbers, and constructs a two-dimensional discourse flow space. The two-dimensional discourse flow space is divided into an ontology feature flow coordinate layer, an object interaction flow coordinate layer, and an environmental noise flow coordinate layer on the vertical axis. Based on the set of clause-level semantic units with timestamp tags and ordinal numbers and the two-dimensional discourse flow space, a two-dimensional discourse flow space with cross-layer dependency pointers is obtained. Based on a two-dimensional discourse flow space with cross-layer dependency pointers, the density parameters of medical entities in the object interaction flow coordinate layer of the two-dimensional discourse flow space are calculated. Based on the density parameters of medical entities, the time windows that need to be blocked are identified, and a blocking protection buffer distance interval is generated for each time window that needs to be blocked. Within the blocking protection buffer distance interval, the subject-verb-object relationship re-verification processing is performed on the medical entities in the ontology feature flow coordinate layer, and a structured case summary text is generated.

2. The dual-track entity recognition spoken language denoising and parsing method for multi-source heterogeneous long text according to claim 1, characterized in that, The two-layer decomposition includes a first-layer timestamp slicing process and a second-layer sentence-level semantic decomposition process; The first-level timestamp slicing process involves slicing the original text of the patient's daily conversation records according to the message sending timestamp, resulting in a message slice sequence arranged in chronological order. The second-level semantic decomposition process is as follows: for each message slice in the message slice sequence arranged in chronological order, a joint delimitation mechanism of punctuation mark boundaries and transitional conjunction positions is used to decompose it into several independent semantic units at the clause level. Preserve the timestamp labels and ordinal numbers to construct a set of clause-level semantic units with timestamp labels and ordinal numbers.

3. The dual-track entity recognition spoken language denoising and parsing method for multi-source heterogeneous long text according to claim 1, characterized in that, The method for constructing a two-dimensional discourse flow space includes: Based on the set of clause-level semantic units with timestamp labels and ordinal numbers, a set of clause-level semantic units with subject category labels is obtained; The timestamps of message slice sequences arranged in chronological order are extracted as the horizontal axis coordinates, and the subject category labels of the sentence-level semantic unit sets with subject category labels are extracted as the vertical axis coordinates to establish a two-dimensional discourse flow space.

4. The dual-track entity recognition spoken language denoising and parsing method for multi-source heterogeneous long text according to claim 3, characterized in that, The method for generating the set of clause-level semantic units with subject category labels includes: For each clause-level semantic unit in the set of clause-level semantic units with timestamp labels and ordinal numbers, logical subject identification processing is performed. Based on the logical subject identification results, each clause-level semantic unit is assigned one of three subject category labels: ontology feature flow label, object interaction flow label, or environmental noise flow label, resulting in a set of clause-level semantic units with subject category labels.

5. The dual-track entity recognition spoken language denoising and parsing method for multi-source heterogeneous long text according to claim 3, characterized in that, The method for generating the two-dimensional discourse flow space with cross-layer dependency pointers includes: Each clause-level semantic unit in the set of clause-level semantic units with subject category labels is projected to the corresponding coordinate position on the corresponding coordinate layer in the two-dimensional discourse flow space according to the timestamp and subject category label. Cross-layer dependency pointers are established for clause-level semantic units that originate from the same message slice but belong to different coordinate layers, and the two-dimensional discourse flow space with cross-layer dependency pointers is output.

6. The dual-track entity recognition spoken language denoising and parsing method for multi-source heterogeneous long text according to claim 1, characterized in that, The calculation method for the density parameter of the medical entity includes: Medical entity scanning processing is performed on all clause-level semantic units in the object interaction flow coordinate layer of the two-dimensional discourse flow transition space with cross-layer dependency pointers to obtain a list of medical entities in the object interaction flow. The object interaction flow coordinate layer is divided into continuous equal-width time windows according to the horizontal axis time dimension. The ratio of the number of medical entities recorded in the object interaction flow medical entity list to the total number of clause-level semantic units in each equal-width time window is calculated to obtain the medical entity density parameter for each equal-width time window.

7. The dual-track entity recognition spoken language denoising and parsing method for multi-source heterogeneous long text according to claim 6, characterized in that, The method for identifying the time window that needs to be blocked includes: Set a shielding trigger threshold parameter and compare the density parameter of medical entities in each time window with the shielding trigger threshold parameter. When the density parameter of medical entities in a certain equal-width time window is greater than or equal to the shielding trigger threshold parameter, this equal-width time window is determined to be a time window that needs to be shielded. Apply a shielding mask mark to the medical entity list of the object interaction flow for the time window that is determined to be a time window that needs to be shielded.

8. The dual-track entity recognition spoken language denoising and parsing method for multi-source heterogeneous long text according to claim 6, characterized in that, The method for generating the structured case summary text includes: From the ontology feature flow coordinate layer data after subject-verb-object relation re-verification processing, all medical entities with valid attribution status labels are selected to obtain the confirmed ontology medical entity set. The confirmed ontology medical entity set is sorted according to the horizontal axis time dimension to form a temporal ontology medical entity sequence. The temporal ontology medical entity sequence is anchored and matched with the discharge summary structured baseline dataset and the ontology feature flow coordinate layer data with baseline anchoring mapping to extract the set of change trend features before and after intervention. The confirmed ontology medical entity set and the set of change trend features before and after intervention are processed to generate a structured case closure summary text.

9. The dual-track entity recognition spoken language denoising and parsing method for multi-source heterogeneous long text according to claim 8, characterized in that, The method for obtaining the ontology feature flow coordinate layer data with baseline anchoring mapping includes: Medical entity pre-scanning is performed on clause-level semantic units in the ontology feature flow coordinate layer of the two-dimensional discourse flow space with cross-layer dependency pointers. The medical entities extracted by the pre-scanning are anchored and mapped with the structured baseline dataset of the discharge summary to obtain ontology feature flow coordinate layer data with baseline anchor mapping. The structured baseline dataset of discharge summaries is obtained by receiving patient discharge summary documents and performing structured parsing processing on the patient discharge summary documents.

10. The dual-track entity recognition spoken language denoising and parsing method for multi-source heterogeneous long text according to claim 9, characterized in that, The anchoring mapping method is as follows: Each medical entity extracted from the pre-scan medical entity is compared with the semantic similarity of each medical entity in the structured baseline dataset of the discharge summary. When the semantic similarity between a medical entity extracted from the pre-scan medical entity and a medical entity in the structured baseline dataset of the discharge summary is greater than or equal to the preset anchor mapping similarity threshold, an anchor mapping link is established between the two.

11. The dual-track entity recognition spoken language denoising and parsing method for multi-source heterogeneous long text according to claim 10, characterized in that, The method for extracting the set of trend features before and after the anchor intervention includes: For each medical entity in the temporally sequenced ontology medical entity sequence, the anchor mapping link corresponding to the medical entity is found in the ontology feature flow coordinate layer data with baseline anchor mapping. The anchor mapping link is used to trace back to the corresponding medical entity in the structured baseline dataset of the discharge summary to obtain the baseline state value at the time of discharge. The baseline state value at the time of discharge is compared with the current state description mentioned in the patient's daily conversations to extract the information on the direction and magnitude of change, forming a set of trend features before and after intervention.

12. A dual-track entity recognition spoken language denoising and parsing system for multi-source heterogeneous long texts, used to implement the dual-track entity recognition spoken language denoising and parsing method for multi-source heterogeneous long texts as described in any one of claims 1-11, characterized in that, The system includes: The decomposition space construction module is used to receive the original text of the patient's daily dialogue record and the patient's discharge summary document, perform two-layer decomposition on the original text of the patient's daily dialogue record, generate a set of clause-level semantic units with timestamp tags and sequence numbers, and construct a two-dimensional discourse flow space. The two-dimensional discourse flow space is divided into an ontology feature flow coordinate layer, an object interaction flow coordinate layer and an environmental noise flow coordinate layer on the vertical axis. Cross-layer pointer construction module: Based on the set of clause-level semantic units with timestamp tags and ordinal numbers and the two-dimensional discourse flow space, a two-dimensional discourse flow space with cross-layer dependency pointers is obtained; The shielding interval generation module: Based on the two-dimensional discourse flow space with cross-layer dependency pointers, it calculates the density parameters of medical entities in the object interaction flow coordinate layer of the two-dimensional discourse flow space, identifies the time windows that need to be shielded based on the density parameters of medical entities, and generates the shielding protection buffer distance interval for each time window that needs to be shielded. Entity Verification and Closure Module: This module is used to perform subject-verb-object re-verification processing on medical entities within the ontology feature flow coordinate layer within the shielded protection buffer distance range, and generate a structured closure summary text.

Images