CN-122025108-A - Method, apparatus and storage medium for retrieving medical information

CN122025108ACN 122025108 ACN122025108 ACN 122025108ACN-122025108-A

Abstract

The present disclosure relates to a method, apparatus and storage medium for retrieving medical information. The method comprises the steps of generating a structural analysis sequence based on original input information input by a user through a pre-trained large language model, wherein the structural analysis sequence at least comprises disease entity text characterization information and query intention information, the query intention information is related to the disease entity text characterization information, encoding through a pre-trained medical language model based on the disease entity text characterization information to obtain a plurality of candidate standard disease entity information, the pre-trained medical language model is a double-tower model sharing encoder weights, reordering the plurality of candidate standard disease entity information to obtain target standard disease items, constructing a structural semantic analysis result based on the target standard disease items and the query intention information to search, and accurately analyzing disease entities and query intents in questions input by the user to remarkably improve the accuracy of the search result.

Inventors

CHEN ZIXIANG
FENG JING
LI YUEZHEN
GAO YANG

Assignees

北京贝瑞和康生物技术有限公司

Dates

Publication Date: 20260512
Application Date: 20260410

Claims (10)

1. A method for retrieving medical information, comprising: based on original input information input by a user, generating a structured parsing sequence through a pre-trained large language model, wherein the structured parsing sequence at least comprises disease entity text characterization information and query intention information, and the query intention information is associated with the disease entity text characterization information; encoding via a pre-trained medical language model based on the disease entity text characterization information to obtain a plurality of candidate standard disease entity information, the pre-trained medical language model being a dual tower model sharing encoder weights, and Reordering the plurality of candidate standard disease entity information to obtain a target standard disease entry; Based on the target standard disease item and the query intent information, a structured semantic parsing result is constructed for retrieval based on the structured semantic parsing result.
2. The method of claim 1, wherein obtaining a plurality of candidate standard disease entity information based on the disease entity text characterization information encoded via a pre-trained medical language model comprises: extracting semantic features of disease entity text characterization information to obtain semantic vectors using a pre-trained medical language model comprising at least an encoder and a bilinear projection network layer, and And performing neighbor retrieval in a standard disease semantic vector library at least based on the semantic vector so as to acquire a plurality of candidate standard disease entity information.
3. The method of claim 2, wherein the twin tower model is constructed based on a KBioXLM model, wherein performing neighbor retrieval in a standard disease semantic vector library based at least on the semantic vectors to obtain a plurality of candidate standard disease entity information comprises: non-linear mapping of semantic vectors to obtain projected semantic vectors, and And based on the projected semantic vector, performing neighbor retrieval in a standard disease semantic vector library so as to acquire a plurality of candidate standard disease entity information.
4. The method of claim 1, wherein generating a structured analysis sequence via a pre-trained large language model based on raw input information entered by a user comprises: Generating input data based on the original input information entered by the user, a predetermined set of query intents including a plurality of query intent normalization information, and structured task instructions, and Features of the input data are extracted via the pre-trained large language model to obtain a generated structured analytical sequence comprising at least disease entity text characterization information and query intent information.
5. The method of claim 2, wherein reordering the plurality of candidate standard disease entity information to obtain the target standard disease entry comprises: And splicing the disease entity text representation information with the plurality of candidate standard disease entity information so as to input a reordering model for reordering the plurality of candidate standard disease entity information, thereby obtaining target standard disease items, wherein the reordering model is of a cross coding structure.
6. The method of claim 5, wherein reordering the plurality of candidate standard disease entity information to obtain the target standard disease entry comprises: Performing correlation calculation for a plurality of candidate standard disease entity information and standard disease entity information, and Based on the correlation score calculation result, reordering is performed for the plurality of candidate standard disease entity information to obtain a target standard disease entry based on the reordered result.
7. A method according to claim 3, wherein the dual tower model is trained via: Constructing a corresponding positive sample aiming at each disease entity text characterization information; Constructing a plurality of negative samples aiming at each disease entity text characterization information; Training is performed for the dual tower model based on the corresponding positive samples, the plurality of negative samples, and the contrast learning loss function.
8. The method of claim 7, wherein the plurality of negative samples comprises a plurality of: randomly selecting a disease item which is not matched with the text characterization information of the current disease entity from a preset database; Selecting a disease item which is similar to the text characterization information name or the semantic of the current disease entity but not the same disease; Disease entries recalled via the double tower model that are arranged in a predetermined number of previous recall results that do not match current disease entity text characterization information, and Disease entries within the same training batch that match other disease entity text characterization information.
9. A computing device, comprising: At least one processing unit; at least one memory coupled to the at least one processing unit and storing instructions for execution by the at least one processing unit, the instructions when executed by the at least one processing unit cause the computing device to perform the method of any of claims 1 to 8.
10. A computer readable storage medium having stored thereon a computer program which, when executed by a machine, implements the method according to any of claims 1 to 8.

Description

Method, apparatus and storage medium for retrieving medical information Technical Field The present disclosure relates generally to information retrieval and artificial intelligence, and in particular, to methods, computing devices, and computer storage media for retrieving medical information. Background With the development of accurate medicine and the continuous deepening of clinical auxiliary decision-making systems, a semantic understanding and question-answering engine or a retrieval system constructed based on structured disease data is a core component in a hospital intelligent system, a medical knowledge service platform and a scientific research support tool. In the traditional method for retrieving medical information, the expression mode of the natural language question input by the user is free, fuzzy and has strong semantic uncertainty. For example, the disease name in the natural language question input by the user may be in chinese, english, abbreviation, alias or historical naming, and there is ambiguity and semantic drift with the disease entry in the standard database, so it is difficult to map directly to the standard database entry, and therefore it is difficult to accurately analyze the disease entity and query intention in the question input by the user, and it is further difficult to respond to the user question accurately. In summary, the conventional method for retrieving medical information has the disadvantages that it is difficult to accurately analyze the disease entity and the query intention in the question inputted by the user, and thus it is difficult to accurately respond to the user question. Disclosure of Invention The present disclosure provides a method, a computing device, and a computer storage medium for retrieving medical information, capable of accurately resolving a disease entity and a query intention in a question inputted by a user, and significantly improving accuracy of a retrieval result. According to a first aspect of the present disclosure, a method for retrieving medical information is provided. The method comprises the steps of generating a structural analysis sequence based on original input information input by a user through a pre-trained large language model, wherein the structural analysis sequence at least comprises disease entity text characterization information and query intention information, the query intention information is associated with the disease entity text characterization information, encoding through a pre-trained medical language model based on the disease entity text characterization information to obtain a plurality of candidate standard disease entity information, the pre-trained medical language model is a double-tower model sharing encoder weights, and reordering the plurality of candidate standard disease entity information to obtain target standard disease items, and constructing a structural semantic analysis result based on the target standard disease items and the query intention information for searching based on the structural semantic analysis result. According to a second aspect of the present invention there is also provided a computing device comprising a memory configured to store one or more computer programs, and a processor coupled to the memory and configured to execute the one or more programs to cause the apparatus to perform the method of the first aspect of the present disclosure. According to a third aspect of the present disclosure, there is also provided a non-transitory computer-readable storage medium. The non-transitory computer-readable storage medium has stored thereon machine-executable instructions that, when executed, cause a machine to perform the method of the first aspect of the present disclosure. In some embodiments, encoding via a pre-trained medical language model based on disease entity text characterization information, obtaining a plurality of candidate standard disease entity information includes extracting semantic features of the disease entity text characterization information using the pre-trained medical language model to obtain semantic vectors, the pre-trained medical language model including at least an encoder and a bilinear projection network layer, and performing neighbor retrieval in a standard disease semantic vector library based at least on the semantic vectors to obtain the plurality of candidate standard disease entity information. In some embodiments, the dual-tower model is constructed based on KBioXLM models, performing a neighbor search in a standard disease semantic vector library based at least on the semantic vector to obtain a plurality of candidate standard disease entity information includes performing a non-linear mapping on the semantic vector to obtain a projected semantic vector, and performing a neighbor search in the standard disease semantic vector library based on the projected semantic vector to obtain a plurality of candidate standard disease entit