AD and MCI knowledge discovery method and device, electronic equipment and medium

Abstract

The invention provides an AD and MCI knowledge discovery method and device, electronic equipment and a medium, and the method comprises the steps: obtaining documents in the AD / MCI field, and constructing an AD / MCI knowledge base based on the documents; obtaining a question input by a user, and pre-judging the question to determine a question answering mode; carrying out AD / MCI knowledge base retrieval and / or AD / MCI large language model interaction based on a question answering mode to obtain a question answering result; wherein the question answering result comprises one of a first final retrieval result, a first interaction result, a second final retrieval result and a second interaction result based on the second final retrieval result; if a second interaction result is obtained, correlation evaluation scoring and consistency evaluation scoring are carried out on the second interaction result, and if the score of the second interaction result meets a preset condition, the second interaction result is fed back to the user. According to the invention, the efficiency and quality of obtaining and discovering AD / MCI related knowledge are improved.

Description

Technical Field

[0001] The present invention relates to the field of neuroscience technology, and in particular to an AD and MCI knowledge discovery method, device, electronic equipment and medium. Background Art

[0002] Currently, with the gradual deepening of aging, chronic diseases such as Alzheimer's disease (AD) and mild cognitive impairment (MCI) have brought heavy burdens to patients, families and society; but at the same time, there is still a lack of understanding of the risk factors, mechanisms and intervention methods of AD and mild cognitive impairment.

[0003] In order to better deal with Alzheimer's disease, on the one hand, more medical research needs to be carried out to study the relevant important factors; on the other hand, considering that the research results are mainly published through academic literature, how to efficiently discover relevant literature and knowledge is a problem; especially, with the rapid growth of academic literature in recent years, considering the time constraints of researchers in the AD / MCI field, how to efficiently discover relevant literature and how to efficiently obtain relevant field progress knowledge has become an extremely important issue.

[0004] In this regard, the most important means for researchers in the field to obtain literature and knowledge of field progress is mainly through manual acquisition and reading of literature, which directly leads to problems such as time-consuming, low efficiency and low quality. Although there are some auxiliary tools at present, they are still very insufficient in terms of convenience, accuracy and efficiency. Therefore, in the context of knowledge explosion, the low efficiency and low quality of obtaining and discovering AD / MCI-related literature, evidence, knowledge, ideas, etc. has become an urgent problem to be solved. Summary of the invention

[0005] In view of this, the purpose of the present invention is to provide an AD and MCI knowledge discovery method, device, electronic device and medium to improve the efficiency and quality of acquiring and discovering AD / MCI related knowledge, reduce workload and complexity, and improve the effect of knowledge discovery.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0007] In a first aspect, the present invention provides an AD and MCI knowledge discovery method, comprising: obtaining literature in the AD / MCI field, and constructing an AD / MCI knowledge base based on the literature; wherein the AD / MCI knowledge base includes: an AD / MCI literature base, an AD / MCI literature entity base, an AD / MCI evidence base, and an AD / MCI index base; obtaining a question input by a user, and pre-judging the question to determine a way to answer the question; performing an AD / MCI knowledge base search and / or AD / MCI large language model interaction based on the way to answer the question, to obtain a question answering result; wherein the question answering result includes one of the following: a first final retrieval result, a first interaction result, a second final retrieval result, and a second interaction result based on the second final retrieval result; if the first final retrieval result or the first interaction result is obtained, the first final retrieval result or the first interaction result is fed back to the user, and if the second interaction result is obtained, the second interaction result is evaluated and scored for relevance and consistency, and if the score of the second interaction result meets a preset condition, the second interaction result is fed back to the user.

[0008] Optionally, constructing an AD / MCI knowledge base based on documents includes: extracting metadata of each document according to the labels of the knowledge system of the pre-constructed AD / MCI document library, and processing the metadata based on the pre-constructed knowledge rules, classification models and information extraction models to obtain the content corresponding to each label in the knowledge system, and constructing an AD / MCI document library according to the content and metadata corresponding to each label of each document; wherein the metadata also includes the document structure and content; for each document in the AD / MCI document library, extracting entities in each document based on the pre-constructed entity information extraction model to construct an AD / MCI document entity library; for each document in the AD / MCI document library, if the document If the document type tag content includes RCT, the evidence information in each document is extracted based on the pre-built evidence extraction model to build an AD / MCI evidence library; wherein the evidence information is information related to RCT; sub-indexes of the AD / MCI document library, AD / MCI document entity library and AD / MCI evidence library are respectively constructed, and based on the association ID between the data of the AD / MCI document library, AD / MCI document entity library and AD / MCI evidence library, the AD / MCI document library, AD / MCI document entity library and AD / MCI evidence library are integrated and associated to construct an overall index to obtain the AD / MCI index library; wherein the sub-index and the overall index include an inverted index and a vector index.

[0009] Optionally, a method of answering the question is determined by pre-judgment of the question, including: objectivity judgment, AD / MCI domain judgment and answerability judgment; if the question is an objective question, a question in the AD / MCI domain, or an answerable question, the method of answering the question is determined based on a pre-built question answering method judgment model; among which, the methods of answering questions include: AD / MCI knowledge base retrieval method, AD / MCI large language model interactive answer method, and a combination of AD / MCI knowledge base retrieval and AD / MCI large language model interactive answer method.

[0010] Optionally, AD / MCI knowledge base retrieval and / or AD / MCI large language model interaction are performed based on the way of answering the question to obtain the question answer result, including: if the way of answering the question is AD / MCI knowledge base retrieval, then AD / MCI knowledge base retrieval is performed based on the question to obtain a first final retrieval result; if the way of answering the question is AD / MCI large language model interactive answering, then the question is input into the AD / MCI large language model to obtain a first interactive result; if the way of answering the question is a combination of AD / MCI knowledge base retrieval and AD / MCI large language model interactive answering, then the question is preprocessed, and AD / MCI knowledge base retrieval is performed based on the preprocessed question to obtain a second final retrieval result, and a prompt of the AD / MCI large language model is constructed based on the second final retrieval result, and the prompt is input into the AD / MCI large language model to obtain a second interactive result based on the second final retrieval result.

[0011] Optionally, the question is preprocessed, and an AD / MCI knowledge base search is performed based on the preprocessed question to obtain a second final search result, including: inputting the question into an AD / MCI large language model to decompose the question, and splicing the decomposition result with the question to obtain a spliced ​​query question; performing word segmentation and keyword extraction on the spliced ​​query question to obtain a first keyword list, and inputting the first keyword list into a pre-built synonym large language model to obtain a first synonym list of the keyword; extracting entities of the question based on a pre-built information extraction model to obtain a first entity list; performing search condition judgment on the question based on a pre-built judgment model to obtain a judgment condition result; wherein the search condition includes at least: the document type to which the target content belongs, and the position of the target content in the document; determining a first search logic for an inverted index of the AD / MCI document library based on the first synonym list of keywords and the judgment condition result, and determining a first search logic for an inverted index of the AD / MCI document library based on the spliced ​​query question and the judgment condition result. A second retrieval logic for the vector index of the AD / MCI document library is determined, and an inverted index retrieval is performed based on the first retrieval logic to obtain a first document retrieval result, and a vector retrieval is performed based on the second retrieval logic to obtain a second document retrieval result; a third retrieval logic for the AD / MCI document entity library is determined based on the first entity list and the judgment condition result, and the AD / MCI document entity library is searched based on the third retrieval logic to obtain a first entity retrieval result; if the document type in the judgment condition result includes RCT, a fourth retrieval logic for the AD / MCI evidence library is determined based on the first entity list and the first synonym list of keywords, and the AD / MCI evidence library is searched based on the fourth retrieval logic to obtain a first evidence retrieval result; the question, the first document retrieval result, the second document retrieval result, the first entity retrieval result and the first evidence retrieval result are input into a pre-constructed first evaluation scoring model to obtain a scoring result for each retrieval result, and a second final retrieval result is determined based on the scoring result.

[0012] Optionally, an AD / MCI knowledge base search is performed based on the question to obtain a first final search result, including: performing word segmentation and keyword extraction on the question to obtain a second keyword list, and inputting the second keyword list into a pre-built synonym large language model to obtain a second synonym list for the keywords; extracting entities of the question based on a pre-built information extraction model to obtain a second entity list; constructing a fifth search logic for the AD / MCI document library, a sixth search logic for the AD / MCI document entity library, and a seventh search logic for the AD / MCI evidence library based on the second synonym list and the second entity list of the keywords; searching the AD / MCI document library based on the fifth search logic to obtain a third document search result, searching the AD / MCI document entity library based on the sixth search logic to obtain a second entity search result, and searching the AD / MCI evidence library based on the seventh search logic to obtain a second evidence search result; inputting the question, the third document search result, the second entity search result, and the second evidence search result into a pre-built first evaluation scoring model to obtain a scoring result for each search result, and determining the first final search result based on the scoring results.

[0013] Optionally, constructing a prompt for an AD / MCI large language model based on the second final search result includes: determining the role, tasks and requirements of the AD / MCI large language model; after preprocessing the second final search result, sorting the preprocessed second final search result based on the scoring result of the second final search result to obtain reference materials; wherein the preprocessing of the second final search result includes: summarizing and compressing the document content in the second final search result, compressing the content in the second final search result that exceeds a preset length, and combining the evidence content in the second final search result into a natural language sentence; determining clue information based on the first keyword list and the first entity list; constructing a prompt for an AD / MCI large language model based on the question, the role, tasks, requirements, reference materials and clue information of the AD / MCI large language model.

[0014] Optionally, the second interaction result is evaluated and scored for relevance and consistency, and if the score of the second interaction result meets preset conditions, the second interaction result is fed back to the user, including: filtering the second interaction result based on pre-set rules; inputting the filtered second interaction result into a pre-constructed second evaluation and scoring model to obtain an evaluation score of the second interaction result; if the evaluation score of the second interaction result is greater than or equal to a first threshold, the second interaction result is fed back to the user; if the evaluation score of the second interaction result is greater than or equal to the second threshold and less than the first threshold, the second interaction result is evaluated and scored for consistency; wherein the first threshold is greater than the second threshold; if the evaluation score of the second interaction result is less than the second threshold, the prompt of the AD / MCI large language model is reconstructed to obtain a new second interaction result, and the new second interaction result is evaluated and scored for relevance; if the evaluation score of the new second interaction result is less than the second threshold, the question is re-preprocessed, and the pre-processed question is searched through the AD / MCI knowledge base and interacted with the AD / MCI large language model to obtain a new second interaction result.

[0015] Optionally, the second interaction result is evaluated and scored for consistency, including: constructing a key summary prompt based on the second interaction result, and inputting the key summary prompt into the AD / MCI large language model to obtain the key summary result of the second interaction result; scoring the key summary result based on a pre-constructed third evaluation and scoring model to obtain an evaluation score of the key summary result; if the evaluation score of the key summary result is greater than or equal to a third threshold, the second interaction result is fed back to the user; if the evaluation score of the key summary result is less than the third threshold, the prompt of the AD / MCI large language model is reconstructed to obtain a new second interaction result, and the key points of the new second interaction result are summarized and scored. If the evaluation score of the key point summary result of the new second interaction result is less than the third threshold, the question is pre-processed again, and the pre-processed question is searched through the AD / MCI knowledge base and interacted with the AD / MCI large language model to obtain a new second interaction result.

[0016] In a second aspect, the present invention provides an AD and MCI knowledge discovery device, including: a knowledge base construction module, used to obtain literature in the AD / MCI field and construct an AD / MCI knowledge base based on the literature; wherein the AD / MCI knowledge base includes: an AD / MCI literature library, an AD / MCI literature entity library, an AD / MCI evidence library, and an AD / MCI index library; a pre-judgment module, used to obtain questions input by users, and pre-judgment the questions to determine the way to answer the questions; a retrieval and interaction module, used to perform AD / MCI knowledge base retrieval and / or AD / MCI large language model interaction based on the way to answer the questions, and obtain question answering results; wherein the question answering results include one of the following: a first final retrieval result, a first interaction result, a second final retrieval result, and a second interaction result based on the second final retrieval result; a result evaluation and feedback module, used to feed back the first final retrieval result or the first interaction result to the user if the first final retrieval result or the first interaction result is obtained, and if the second interaction result is obtained, the second interaction result is evaluated and scored for relevance and consistency, and if the score of the second interaction result meets the preset conditions, the second interaction result is fed back to the user.

[0017] In a third aspect, the present invention provides an electronic device comprising a processor and a memory, wherein the memory stores computer-executable instructions that can be executed by the processor, and the processor executes the computer-executable instructions to implement the steps of any one of the methods provided in the first aspect above.

[0018] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of any one of the methods provided in the first aspect are executed.

[0019] The present invention brings the following beneficial effects:

[0020] The above-mentioned AD and MCI knowledge discovery method, device, electronic device and medium provided by the present invention first obtain documents in the AD / MCI field, and build an AD / MCI knowledge base (AD / MCI document base, AD / MCI document entity base, AD / MCI evidence base, AD / MCI index base) based on the documents; then obtain the question input by the user, and pre-judge the question to determine the way to answer the question; then, based on the way to answer the question, perform AD / MCI knowledge base retrieval and / or AD / MCI large language model interaction to obtain the question answering result (the question answering result includes one of the following: a first final retrieval result, a first interaction result, a second final retrieval result and a second interaction result based on the second final retrieval result); finally, if the first final retrieval result or the first interaction result is obtained, the first final retrieval result or the first interaction result is fed back to the user; if the second interaction result is obtained, the second interaction result is scored for relevance evaluation and consistency evaluation; if the score of the second interaction result meets the preset conditions, the second interaction result is fed back to the user. The above method ensures that AD / MCI knowledge is comprehensive and up-to-date by constructing a real-time updated, multi-dimensional AD / MCI knowledge base including documents, entities, evidence, indexes, etc.; by pre-judging the user's questions and matching the complexity of the questions with the answering methods, the accuracy and effectiveness are guaranteed on the basis of ensuring the efficiency of knowledge discovery; by performing relevance evaluation and consistency evaluation on the retrieval results of the user's questions and the interaction results of the AD / MCI large language model, the accuracy of the response results is guaranteed, thereby improving the efficiency and quality of obtaining and discovering AD / MCI related knowledge and reducing the workload and complexity.

[0021] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0022] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the preferred embodiments are specifically mentioned below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0024] Figure 1A flowchart of an AD and MCI knowledge discovery method provided by an embodiment of the present invention;

[0025] Figure 2 A schematic diagram of the structure of an AD and MCI knowledge discovery device provided by an embodiment of the present invention;

[0026] Figure 3 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0027] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0028] At present, the most important means for researchers in the field to obtain literature and knowledge of field progress is mainly through manual acquisition and manual reading of literature, etc.; some existing auxiliary tools still have great deficiencies in terms of convenience, accuracy and efficiency. Therefore, in the context of knowledge explosion, the low efficiency and low quality of obtaining and discovering AD / MCI-related literature, evidence, knowledge, ideas, etc. have become a problem that needs to be solved urgently.

[0029] Based on this, an AD and MCI knowledge discovery method, device, electronic device and medium provided by the embodiments of the present invention can improve the efficiency and quality of acquiring and discovering AD / MCI related knowledge, reduce workload and complexity, and improve the effect of knowledge discovery.

[0030] To facilitate understanding of this embodiment, a method for discovering AD and MCI knowledge disclosed in an embodiment of the present invention is first introduced in detail. The method can be executed by an electronic device, such as a smart phone, a computer, a tablet computer, etc. Figure 1 The flowchart of a method for discovering AD and MCI knowledge is shown, which illustrates that the method mainly includes the following steps S101 to S104:

[0031] Step S101: Acquire literature in the field of AD / MCI, and build an AD / MCI knowledge base based on the literature.

[0032] In one implementation, in order to support AD / MCI-oriented intelligent knowledge discovery, it is necessary to build a knowledge base of AD / MCI-related research results. To this end, it is necessary to first build an AD / MCI knowledge base framework. Specifically, the AD / MCI knowledge base includes: AD / MCI document base, AD / MCI document entity base, AD / MCI evidence base, and AD / MCI index base.

[0033] In order to ensure the completeness of the AD / MCI knowledge base, AD / MCI literature can be retrieved from multiple databases, such as pubmed, embase, cochrane, Wanfang Medical, etc. For the obtained AD / MCI literature, machine learning methods are further used to distinguish whether the literature belongs to AD / MCI. After that, the retained literature is processed by extracting metadata, processing based on rules (pre-set), building a classification model, and building an information extraction model to form a complete AD / MCI literature library; and based on the constructed entity information extraction model, an AD / MCI literature entity library is formed; for the literature with RCT label content in its document type, an AD / MCI evidence library is formed based on the constructed evidence extraction model; finally, an AD / MCI index library is constructed.

[0034] Step S102: Obtain the question input by the user, and pre-judge the question to determine the method of answering the question.

[0035] In one embodiment, the user service interface includes traditional keyword input, and users can input related questions by inputting keywords, or express their AD / MCI related questions in natural language.

[0036] In order to ensure the quality of knowledge discovery, in this embodiment, a pre-judgment of the question Query input by the user can be performed to determine the method of answering the question, namely, the AD / MCI knowledge base retrieval method, the AD / MCI large language model interactive answer method, or the combination of AD / MCI knowledge base retrieval and AD / MCI large language model interactive answer method, and corresponding processing can be performed based on the judgment result.

[0037] Step S103: Perform AD / MCI knowledge base search and / or AD / MCI large language model interaction based on the question answering method to obtain the question answering result.

[0038] In one embodiment, based on the above-mentioned question and the determined way to answer the question, an AD / MCI knowledge base search and / or an AD / MCI large language model interaction is performed to obtain a question answering result; wherein the question answering result includes one of the following: a first final search result, a first interaction result, a second final search result, and a second interaction result based on the second final search result. The AD / MCI knowledge base search includes: a search for an AD / MCI document library, a search for an AD / MCI entity library, and a search for an AD / MCI evidence library.

[0039] Step S104: If the first final retrieval result or the first interaction result is obtained, the first final retrieval result or the first interaction result is fed back to the user. If the second interaction result is obtained, the second interaction result is scored for relevance and consistency. If the score of the second interaction result meets the preset conditions, the second interaction result is fed back to the user.

[0040] In one embodiment, if the AD / MCI knowledge base search obtains the first final search result, the first final search result is directly fed back to the user; if the AD / MCI large language model interaction obtains the first interaction result, the first interaction result is directly fed back to the user; if the AD / MCI knowledge base search and the AD / MCI large language model interaction obtain the second final search result and the second interaction result based on the second final search result, then after obtaining the second interaction result of the AD / MCI large language model, the obtained second interaction result can be checked and filtered for content, and evaluated and scored for relevance. Afterwards, based on the obtained second interaction result Re for the question Query input by the user, a key summary prompt prompt is constructed and sent to the AD / MCI large language model to generate a key summary result Re for the interaction result. sum , and give a consistency evaluation score to the key summary results.

[0041] For the second interaction result Re that has passed the evaluation and scoring (i.e., the score meets the preset conditions, such as the score exceeds the threshold), a reference mark is generated in the second interaction result according to each sentence in the recorded second interaction result and the corresponding reference material number (i.e., document ID), and the reference material content and reference material number are listed after the second interaction result Re, thereby generating the final second interaction result Re last , to provide feedback to users.

[0042] The above-mentioned AD and MCI knowledge discovery method provided by the present invention ensures that AD / MCI knowledge is comprehensive and up-to-date by constructing a real-time updated, multi-dimensional AD / MCI knowledge base including documents, entities, evidence, indexes, etc.; by pre-judging the user's questions and matching the complexity of the questions with the answering methods, it ensures the accuracy and effectiveness while ensuring the efficiency of knowledge discovery; by performing relevance evaluation and consistency evaluation on the retrieval results of the user's questions and the interaction results of the AD / MCI large language model, it ensures the accuracy of the response results, thereby improving the efficiency and quality of acquiring and discovering AD / MCI related knowledge and reducing the workload and complexity.

[0043] In one embodiment, for the aforementioned step S101, that is, when constructing an AD / MCI knowledge base based on literature, the following methods may be used, including but not limited to:

[0044] First, according to the labels of the knowledge system of the pre-constructed AD / MCI document library, the metadata of each document is extracted, and the metadata is processed based on the pre-constructed knowledge rules, classification model and information extraction model to obtain the content corresponding to each label in the knowledge system, and the AD / MCI document library is constructed according to the content and metadata corresponding to each label of each document; wherein the metadata includes title, abstract, keywords, author, author institution, publication date, journal name (and its impact factor), subject classification, number of citations, related documents, etc., as well as document structure and content.

[0045] In specific implementation, the AD / MCI document library framework design can be constructed using a multi-level and multi-link approach. Overall, the first-level categories of the AD / MCI document library framework (the first-level categories of the knowledge system labels) include: document title information, content location information, document type information, disease information, focus information, link information, measurement method information, intervention classification information, intervention information, evaluation standard information, evaluation indicator information, quality evaluation information, methodology evaluation information, comprehensive evaluation information, etc. Further, the first-level categories can be subdivided. For example, the document title information can be divided into: title, abstract, keywords, author, affiliated institution, document publication medium name (journal), impact factor, publication time, etc.; the content location can be further subdivided into: title, abstract, keywords, introduction, background, method, key, experiment, discussion, prospect, etc.; the document type information can be further subdivided into: clinical trials, RCT (Randomized Controlled Trial, randomized controlled trial), meta-analysis, systematic review, review, guideline, others, etc.; disease information can be further subdivided into: AD, MCI, etc.; focus information can be further subdivided into: calculation, reasoning, problem solving, decision-making, perception, memory, attention, visual space, execution, learning, language, etc.; measurement method information can be further subdivided into imaging, EEG, MEG, scale, etc.; link information can be further subdivided into: screening, diagnosis, treatment, prognosis, etc.; intervention classification information can be further subdivided into: Chinese medicine, Western medicine, surgery, non-drug methods, etc.; intervention information can be further subdivided into: Yizhi Jiannao Granules, Yizhi Drink, Dihuang Yizhi Recipe, etc. (taking Chinese medicine intervention as an example): evaluation standard information can be further subdivided into: MCI diagnostic criteria, dementia efficacy evaluation form, etc.; evaluation indicator information can be further subdivided into: MMSE (Mini-mental State Examination, Mini-mental State Examination), MoCA (Montreal Cognitive Assessment), etc.; report quality evaluation can be further divided into: experimental design, data analysis, data interpretation methods and other dimensions; methodological evaluation can be further divided into: random sequence generation method, whether the allocation scheme is hidden, implementation bias, measurement bias, follow-up bias, etc.; comprehensive evaluation information can be further divided into: quality evaluation, recommendation strength, etc. (taking grade as an example). Moreover, the above categories can be further subdivided according to the situation.

[0046] For the AD / MCI documents obtained, since the documents obtained by retrieval and other methods often have a gap with the actual needs, that is, there may be unnecessary documents in the obtained documents, that is, non-AD / MCI documents. Therefore, it is necessary to filter the obtained documents, specifically targeting the title, abstract, keywords, journal name, etc. of the document, using machine learning methods, such as building a model based on the BERT+ tuning method, and distinguishing whether the document belongs to the AD / MCI field based on the constructed model, so as to retain the AD / MCI documents and obtain an accurate AD / MCI document library.

[0047] Furthermore, according to the knowledge system architecture of the above-mentioned AD / MCI document library, each document in the document library is processed by extracting metadata, processing based on knowledge rules (pre-set), building a classification model for processing, and building an information extraction model for processing, and the corresponding label content of each component in the system is obtained and saved in the corresponding fields in the knowledge base (including the paragraphs and structural positions, sentences and specific character positions, etc.), thereby forming a complete AD / MCI document library.

[0048] Secondly, for each document in the AD / MCI document library, entities in each document are extracted based on a pre-built entity information extraction model to build an AD / MCI document entity library.

[0049] In specific implementation, the entity categories of the AD / MCI literature entity library include at least: diseases, symptoms, examinations, tests, operations, scales, brain regions, drugs, prescriptions, Chinese patent medicines, Chinese herbal medicines, surgeries, adverse reactions, data objects, processing algorithms and other categories.

[0050] For each document in the AD / MCI document library, entities in the document are extracted based on a pre-built entity information extraction model (such as an entity extraction model based on seq2seq), and then saved according to fields such as document ID, entity ID, entity category, specific entity, location (including: paragraph and structural position, sentence and specific character position, etc.), and related document data (including the document library knowledge system label and content corresponding to the document) to form an AD / MCI document entity library.

[0051] Then, for each document in the AD / MCI document library, if the document type label content of the document includes RCT, the evidence information in each document is extracted based on the pre-built evidence extraction model to build the AD / MCI evidence library; wherein the evidence information is information related to RCT.

[0052] In specific implementation, the AD / MCI evidence base includes evidence information related to RCT. The specific evidence system is as follows: patient status, diagnostic criteria (such as DSM-IV diagnostic criteria, etc.), grouping situation, grouping basis (such as random number table method, computer random number method, odd-even numbering method, etc.), sample size, demographic data (such as age distribution, gender distribution, etc.), intervention measures (such as placebo treatment, Western medicine treatment, Chinese medicine treatment, etc.), implementation methods (such as type of medicine, dosage, course of treatment), evaluation indicators, group intervention results, adverse events, experimental results, experimental significance, bias risk assessment, etc.

[0053] For each document in the AD / MCI document library, we first determine whether there is RCT tag content in its document type. For documents with RCT tag content, we extract evidence information based on a pre-built evidence extraction model (built in advance, such as an evidence extraction model based on seq2seq), and then save it according to the document ID, evidence ID, evidence field, evidence content, location (including: paragraph and structural position, sentence and specific character position, etc.), related document data (including the document library knowledge system label and content corresponding to the document) and other fields to form an AD / MCI evidence library.

[0054] Finally, sub-indexes of the AD / MCI document library, AD / MCI document entity library and AD / MCI evidence library are constructed respectively, and the AD / MCI document library, AD / MCI document entity library and AD / MCI evidence library are integrated and associated based on the association IDs between the data of the AD / MCI document library, AD / MCI document entity library and AD / MCI evidence library to construct an overall index to obtain the AD / MCI index library; among which, the sub-index and the overall index include an inverted index and a vector index.

[0055] In the specific implementation, when constructing the AD / MCI index library, on the one hand, sub-indexes are constructed for the AD / MCI document library, AD / MCI document entity library and AD / MCI evidence library respectively, and then the data of each library are associated based on the associated ID (such as document ID) between the data of each library to construct an overall index for the entire knowledge base.

[0056] On the other hand, for each sub-index and the overall index, an inverted index (including steps such as word segmentation, and the index field also includes the page number and paragraph number of the text) and a vector index are constructed. In the process of constructing a vector index, for long texts, the text needs to be segmented first (for example, segmented according to the length of 500 characters), then vectorized (based on the pre-trained embedding model), and then the vectorized data is stored in the database separately.

[0057] In order to further improve the efficiency and effect of retrieval, in the process of constructing the vector index, for each segmented text block, based on the fine-tuned large model, the prompt method is used to input the entire document (to which the text block belongs) and the segmented text block into the large model to obtain a context summary for the segmented text block, and then the obtained context summary and the segmented text block are connected to obtain the input text for index construction. In addition, for the AD / MCI evidence library, considering that some of the fields therein include numbers, a digital index is also constructed for the AD / MCI evidence library in this embodiment.

[0058] It should be noted that the AD / MCI knowledge base includes various components such as the AD / MCI document base, AD / MCI document entity base, AD / MCI evidence base, and AD / MCI index base, which can be updated in a timely manner based on the acquisition of source document data, thus ensuring the latest content of the entire AD / MCI knowledge base system.

[0059] In one embodiment, for the aforementioned step S102, that is, when pre-judging the question to determine the method of answering the question, the following methods may be used, including but not limited to: judging the objectivity, AD / MCI field and answerability of the question; if the question is an objective question, a question in the AD / MCI field, or an answerable question, then determining the method of answering the question based on a pre-built question answering method judgment model; wherein the methods of answering questions include: AD / MCI knowledge base retrieval method, AD / MCI large language model interactive answer method, and a combination of AD / MCI knowledge base retrieval and AD / MCI large language model interactive answer method.

[0060] In the specific implementation, for the question Query input by the user, the user question is first pre-judged, including: objectivity judgment, AD / MCI field judgment and answerability judgment. Objectivity judgment is to judge whether the question is objective and whether it requires a true and definite response (the opposite questions include subjective questions, questions that have not yet reached a consensus, imaginary questions, etc.); AD / MCI field judgment is to judge whether the question input by the user belongs to the AD / MCI field-related questions; answerability judgment is to judge whether the question can be answered correctly. The above-mentioned objectivity judgment, AD / MCI field judgment and answerability judgment can be completed by a pre-constructed judgment model (the judgment model can be fine-tuned based on a pre-trained model such as T5). For questions that are judged by the judgment model as non-objective questions, non-AD / MCI fields or non-answerable questions, the user is directly prompted and then returns; for questions that are judged by the judgment model as objective questions, questions in the AD / MCI field, and answerable questions, the following processing is continued to determine the way to answer the question.

[0061] Taking into account the time required to search the knowledge base and the possible lack of knowledge when using the AD / MCI large language model alone, the embodiments of the present invention can further judge the relevant question Query input by the user: judge the way to answer the question, that is, to search the AD / MCI knowledge base for the answer (that is, the AD / MCI knowledge base search method), to answer with the AD / MCI large language model alone (that is, the AD / MCI large language model interactive answer method), and to answer with the AD / MCI large language model based on the search of the AD / MCI knowledge base (that is, the combination of AD / MCI knowledge base search and AD / MCI large language model interactive answer method).

[0062] The judgment of the way to answer a question can be achieved by a pre-constructed question answering way judgment model, wherein the question answering way judgment model can be fine-tuned based on a pre-trained model such as T5, and the fine-tuned training corpus is composed of questions that can be answered correctly by searching the AD / MCI knowledge base, questions that can be answered correctly by the AD / MCI large language model, and questions that can be answered correctly by the AD / MCI large language model based on searching the AD / MCI knowledge base.

[0063] In one embodiment, for the aforementioned step S103, i.e., performing AD / MCI knowledge base retrieval and / or AD / MCI large language model interaction based on the question answering method to obtain the question answering result, the following methods may be used, including but not limited to:

[0064] If the way to answer the question is to search the AD / MCI knowledge base, then search the AD / MCI knowledge base based on the question to obtain a first final search result;

[0065] If the way to answer the question is the way of interactive answering by the AD / MCI large language model, the question is input into the AD / MCI large language model to obtain a first interactive result;

[0066] If the method of answering the question is a combination of AD / MCI knowledge base retrieval and AD / MCI large language model interactive answering, the question is preprocessed, and the AD / MCI knowledge base is searched based on the preprocessed question to obtain a second final retrieval result, and a prompt of the AD / MCI large language model is constructed based on the second final retrieval result, and the prompt is input into the AD / MCI large language model to obtain a second interactive result based on the second final retrieval result.

[0067] Specifically, if the answer mode judgment model determines that only the AD / MCI knowledge base needs to be searched to answer correctly, the AD / MCI knowledge base is directly searched and the result is fed back to the user; if the judgment model determines that only the AD / MCI large language model needs to be used to answer correctly, the question is directly sent to the AD / MCI large language model and the result is fed back to the user; and for questions that can be answered correctly by the AD / MCI large language model based on the AD / MCI knowledge base search, the following processing is continued, that is, the question is preprocessed, and the AD / MCI knowledge base is searched based on the preprocessed question to obtain the second final search result, including the following steps 1 to 8:

[0068] Step 1: Input the question into the AD / MCI large language model for question decomposition, and concatenate the decomposition result with the question to obtain the concatenated query question.

[0069] Specifically, the user's question (Query) is input into the AD / MCI large language model, and the initial feedback of the AD / MCI large language model is obtained, that is, the AD / MCI large language model is prompted to decompose the question by requiring a step-by-step problem-solving method to obtain the decomposition result Answer. 1 , and concatenate the two (decomposition results and questions) to obtain the concatenated query question Query con =(Query, Answer 1 ).

[0070] Step 2: Perform word segmentation and keyword extraction on the concatenated query question to obtain a first keyword list, and input the first keyword list into a pre-built synonym large language model to obtain a first synonym list of the keywords.

[0071] Specifically, concatenate the query question Query con Use a word segmenter (such as Jieba) to segment words and extract keywords (such as using the Textrank algorithm) to obtain the first keyword list Query syn , then input the keyword into the fine-tuned synonym large language model (you can use a base model such as llama) to get the first synonym list of the keyword Query synon (Contains the first keyword list).

[0072] Step 3: Extract the entities of the question based on the pre-built information extraction model to obtain the first entity list.

[0073] Specifically, for the question Query input by the user, a pre-established information extraction model (the same as the information extraction model used in the previous step) is used to extract entities to obtain a first entity list Query entity .

[0074] Step 4: Perform search condition judgment on the question based on the pre-built judgment model to obtain the judgment condition result.

[0075] Specifically, for the query input by the user, a pre-established judgment model (which can be based on a pre-trained model such as T5) is used to perform search condition judgment (i.e., to judge the possible relevant conditions of the query input by the user) and obtain the judgment condition result Query. condition Among them, the search conditions at least include: the type of document to which the target content may belong, the possible location of the target content in the document, such as possible locations: abstract, keywords, introduction, background, methods, key, experiments, discussion, outlook, or possible document types: clinical trials, RCTs, meta-analysis, systematic reviews, reviews, guidelines, others, etc.

[0076] Step 5: Determine the first search logic of the inverted index for the AD / MCI document library based on the first synonym list of the keyword and the judgment condition results, and determine the second search logic of the vector index for the AD / MCI document library based on the splicing query question and the judgment condition results, and perform inverted index search based on the first search logic to obtain the first document search result, and perform vector search based on the second search logic to obtain the second document search result.

[0077] Specifically, searching the AD / MCI document library is divided into two aspects: one is searching the inverted index of the AD / MCI document library, and the other is searching the vector index of the AD / MCI document library.

[0078] Search the inverted index of the AD / MCI document library, and input the first synonym list of the above-processed keywords. synon And the judgment condition result Query condition In order to ensure the comprehensiveness and accuracy of the query, the keyword list (Query synon ) is logically OR, the logical relationship between the judgment conditions is logical OR, and the logical relationship between the keyword list and the judgment condition is logical AND. The final search logic relationship expression (i.e., the first search logic) is And(Or(Query synon ), Or(Query condition )) to get the topK (pre-set, such as 10) search results Search doc-inv (That is, select the first K results in the ranking as the first document search result).

[0079] Search the vector index of the AD / MCI document library, and input the query question Query after the above processing.con And the judgment condition result Query condition , the logical relationship between the judgment conditions is logical OR, and the logical relationship between the concatenated query question and the judgment condition is logical AND, so the final search logic relationship expression (i.e., the second search logic) is And(Query con ,Or(Query condition )) to get the topK (pre-set, such as 10) search results Search doc-vec (That is, select the first K results in the ranking as the second document search results).

[0080] Step 6: Determine the third search logic for the AD / MCI document entity library based on the first entity list and the judgment condition result, and search the AD / MCI document entity library based on the third search logic to obtain the first entity search result.

[0081] Specifically, the AD / MCI entity database is searched, and the input is the first entity list Query entered by the user after the above processing entity , Query of judgment condition results condition , the logical relationship between entity lists is logical OR, the logical relationship between judgment conditions is logical OR, and the logical relationship between entity lists and judgment conditions is logical AND. The final search logic relationship expression (i.e., the third search logic) is And(Or(Query entity ), Or(Query condition )) to get the topK (pre-set, such as 10) search results Search entity (That is, the first K results in the ranking are selected as the first entity search result, and the first entity search result includes: the document content obtained by associating with the document ID, etc.).

[0082] Step 7: If the document type in the judgment condition result includes RCT, the fourth search logic for the AD / MCI evidence base is determined based on the first entity list and the first synonym list of keywords, and the AD / MCI evidence base is searched based on the fourth search logic to obtain the first evidence retrieval result.

[0083] Specifically, in the judgment condition result Query condition In the example, if the judgment result for RCT in the document type information is true (indicated by 1), that is, the document type in the judgment condition result includes RCT, then the AD / MCI evidence database is searched, and the input is the first entity list Query entered by the user after the above processing. entity 、Keyword first synonym list Query synon, the logical relationship between entity lists is logical OR, the logical relationship between keyword lists is logical OR, and the logical relationship between entity list query and keyword list query is logical OR. The final search logic relationship expression (i.e., the fourth search logic) is Or(Or(Query entity ), Or(Query synon )) to get the topK (pre-set, such as 10) search results Search evidence (That is, the first K results in the sorting are selected as the first evidence retrieval result, and the first evidence retrieval result includes: the results composed of the names and contents of each relevant field in the AD / MCI evidence library, etc.).

[0084] Step 8: Input the question, the first document search result, the second document search result, the first entity search result and the first evidence search result into the pre-built first evaluation scoring model to obtain the scoring result of each search result, and determine the second final search result based on the scoring result.

[0085] Specifically, for each search list obtained by the above search, each search result is evaluated and scored by using the pre-constructed first evaluation and scoring model, wherein the first evaluation and scoring model consists of two sub-models Model q 、Model a The two models can have the same structure (such as siamese, etc.), but the training corpus is different. Each training corpus of the former is composed of (question 1: question 2) (the semantics of the two questions in the positive sample are similar, while those in the negative sample are different), and each training corpus of the latter is composed of (question: answer) (the content of the question and answer in the positive sample is matched, while that in the negative sample is not matched). The training goal of the former is to maximize the similarity index of similar (positive sample) questions, and the training goal of the latter is to maximize the correlation index between the positive question and its answer.

[0086] The input of the first evaluation scoring model is the question Query input by the user, each of the above search results (including the first document search result, the second document search result, the first entity search result and the first evidence search result), and the output is the scoring results Score q (i.e. sub-model Model q Scoring results), Score a (i.e. sub-model Model a The scoring result of the search result is Score = W. q *Score q +W a *Score a (Weight W q and Wa can be determined in advance).

[0087] For each search result, the search results can be sorted from large to small according to the above scoring results, and the search results with a score greater than or equal to the preset threshold L can be selected. min The top L (pre-set) results (including duplicate removal, etc.) are used as the second final search result Search last If the second final search result is Search last If the result is empty, step S103 is repeated. If the result is empty again, the user is directly fed back a "cannot answer" result.

[0088] In another embodiment, in the case of directly performing a search to obtain a result, that is, the method of answering the question is a method of searching the AD / MCI knowledge base, when searching the AD / MCI knowledge base based on the question to obtain the first final search result, the following methods may be used, including but not limited to:

[0089] First, the question is segmented and keywords are extracted to obtain a second keyword list, and the second keyword list is input into a pre-built synonym large language model to obtain a second synonym list of the keywords.

[0090] Specifically, the question Query input by the user is segmented by a word segmenter (such as Jieba), keywords are extracted (such as using the Textrank algorithm) to obtain a second keyword list, and a second synonym list Query of the keyword is obtained in a similar manner to the above step 2. synon1 (Contains the second keyword list).

[0091] Secondly, the entities of the question are extracted based on the pre-built information extraction model to obtain a second entity list.

[0092] Specifically, for the question Query input by the user, a pre-established information extraction model (similar to step 3) is used to extract entities to obtain the second entity list Query entity1 .

[0093] Next, based on the second synonym list and the second entity list of keywords, the fifth retrieval logic for the AD / MCI document library (based on the second synonym list of keywords), the sixth retrieval logic for the AD / MCI document entity library (based on the second entity list), and the seventh retrieval logic for the AD / MCI evidence library (based on the second synonym list and the second entity list of keywords) are constructed.

[0094] Then, based on the fifth retrieval logic, the AD / MCI document library is searched to obtain the third document retrieval result, based on the sixth retrieval logic, the AD / MCI document entity library is searched to obtain the second entity retrieval result, and based on the seventh retrieval logic, the AD / MCI evidence library is searched to obtain the second evidence retrieval result.

[0095] Finally, the question, the third document retrieval result, the second entity retrieval result and the second evidence retrieval result are input into a pre-constructed first evaluation scoring model to obtain a scoring result for each retrieval result, and the first final retrieval result is determined based on the scoring result.

[0096] In this embodiment, the specific implementation method of constructing the search logic, searching the AD / MCI document library, the AD / MCI document entity library and the AD / MCI evidence library, merging and sorting the search results, and taking the top M (pre-set) results (including deduplication, etc.) and returning them to the user is similar to the previous embodiment and will not be repeated here.

[0097] In one embodiment, the second final search result Search based on the search in step S103 last , construct a prompt for interacting with the AD / MCI large language model. Specifically, the content and construction process of the prompt include: roles, tasks, requirements, reference materials, clue information, etc. Further, the constructed prompt (including relevant reference materials) is sent to the set AD / MCI large language model, and the second interaction result Re returned by the model is obtained. When constructing the prompt of the AD / MCI large language model based on the second final search result, the following methods can be used, including but not limited to:

[0098] (1) Determine the roles, tasks, and requirements of the AD / MCI large language model.

[0099] In the specific implementation, first clarify the role of the AD / MCI large language model, for example: you are an expert in the field of AD / MCI. Secondly, clarify the tasks of this dialogue, including: listing the questions input by the user (including problem decomposition), and clarify the tasks, such as: requiring AD / MCI knowledge understanding and response based on the user's questions and related materials. Finally, clarify the relevant requirements, such as: the word limit for the response, requiring a realistic answer - if you don't know, just reply "I don't know", etc.

[0100] (2) After preprocessing the second final search results, the preprocessed second final search results are sorted based on the scoring results of the second final search results to obtain reference materials.

[0101] In specific implementation, preprocessing of the second final search result includes: summarizing and compressing the document content in the second final search result, compressing the content in the second final search result that exceeds a preset length, and combining the evidence content in the second final search result into natural language sentences.

[0102] Specifically, for the second final search result Search last If the search result corresponds to a document, the document content needs to be summarized according to the abstract extraction algorithm (such as cluster-based abstract extraction) to compress the content length so that it does not exceed the limit (the allowed length of each reference (i.e., the preset length) is obtained based on the total length of the allowed references / the number of references). doc ); if the search result corresponds to the pre-segmented content, if the content length does not exceed the above allowed length Length doc , no processing is performed, otherwise compression processing is performed; if the search result corresponds to the content of the evidence library, the evidence content is combined into a natural language sentence in the manner of (evidence field: evidence content, ...).

[0103] Then, the search results are arranged in descending order according to their scores, forming a reference list that serves the prompt.

[0104] (3) Determine clue information based on the first keyword list and the first entity list.

[0105] In the specific implementation, list the relevant clue information (such as keywords and entities related to the user's query), and clarify the first keyword list Query obtained in the previous processing process syn , and the first entity list Query entity wait.

[0106] (4) Construct prompts for the AD / MCI large language model based on the question, the roles, tasks, requirements, references, and clue information of the AD / MCI large language model.

[0107] In the specific implementation, the above (1)-(3) are used to form a complete prompt content that interacts with the AD / MCI large language model.

[0108] It should be noted that, in the case where it is judged that "only the AD / MCI large language model is needed to give a correct answer", the prompt content includes the above-mentioned roles, tasks and requirements.

[0109] In one embodiment, for the aforementioned step S104, i.e., when the relevance evaluation and consistency evaluation are performed on the second interaction result, if the score of the second interaction result meets the preset condition, the second interaction result is fed back to the user in the following manners, including but not limited to:

[0110] First, the second interaction result is filtered based on a preset rule.

[0111] In the specific implementation, the content of the second interactive result is checked and filtered. This part of the check and filtering is mainly completed according to pre-set rules. The rules constrain relatively clear knowledge: for example, recommendations for AD / MCI drugs must be consistent with the guidelines, and discriminatory language and results cannot appear, etc. Content that does not comply with the rules will be dealt with accordingly.

[0112] Then, the filtered second interaction result is input into a pre-constructed second evaluation and scoring model to obtain an evaluation and scoring of the second interaction result.

[0113] In the specific implementation, the filtered second interaction result is evaluated and scored, and the evaluation and scoring is completed by the second evaluation and scoring model constructed in advance. In this embodiment, the evaluation and scoring result consists of two parts. On the one hand, it is given by the AD / MCI large language model that generates the second interaction result (by constructing a prompt consisting of roles, tasks-providing a relevance score between 0 and 100, user questions, second interaction results and reference materials, and sending it to the AD / MCI large language model to obtain the relevance score feedback result), including: the relevance score of the second interaction result and the question entered by the user mo -re-qu , and the relevance score between the second interactive result and the search result of the reference included in the prompt mo-re-se (wherein the score is the weighted average of the relevance scores of the second interaction result and each search result), and then according to the pre-set weight (W mo-rq and W mo-rs ) The above two scores are weighted and calculated to obtain the relevance score given by the AD / MCI large language model:

[0114] Relevance model =W mo-rq *Relevance mo-re-qu +W mo-rs *Relevance mo-re-se .

[0115] On the other hand, the relevance score of the second interaction result and the question input by the user is obtained by the pre-trained second evaluation scoring model sub-model (which can be obtained based on the pre-trained model such as T5 tuning). s-re-qu , and the relevance score of the second interactive result and the search result of the reference material listed in the prompt s -re-se (wherein the score is a weighted average of the relevance scores of the second interaction result and each search result), and according to the pre-set weight (W s-rq and W s-rs ) The above two scores are weighted and calculated to obtain the correlation score given by the second evaluation scoring model sub-model:

[0116] Relevance s =W s-rq *Relevance s-re-qu +W s-rs *Relevance s-re-se .

[0117] Finally, we get the evaluation score of the second interaction result for the AD / MCI large language model feedback:

[0118] Relevance=(Relevance s +Relevance model ) / 2 (during the evaluation process, obtain and record each sentence in the second interaction result Re and the corresponding reference material number and content that exceeds the preset threshold).

[0119] Finally, if the evaluation score of the second interaction result is greater than or equal to the first threshold, the second interaction result is fed back to the user; if the evaluation score of the second interaction result is greater than or equal to the second threshold and less than the first threshold, the second interaction result is evaluated and scored for consistency; wherein the first threshold is greater than the second threshold; if the evaluation score of the second interaction result is less than the second threshold, the prompt of the AD / MCI large language model is reconstructed to obtain a new second interaction result, and the new second interaction result is evaluated and scored for relevance; if the evaluation score of the new second interaction result is less than the second threshold, the question is preprocessed again, and the preprocessed question is searched in the AD / MCI knowledge base and interacted with the AD / MCI large language model to obtain a new second interaction result.

[0120] In the specific implementation, the scoring results given by the second evaluation scoring model are processed according to the two pre-set thresholds (the first threshold Re max >Second threshold Re min), if the evaluation score of the second interaction result is greater than or equal to the first threshold Re max , the second interaction result is directly fed back to the user; if the evaluation score of the second interaction result is less than the first threshold Re max But greater than or equal to the second threshold Re min , then the second interaction result is evaluated and scored for consistency; if the evaluation score of the second interaction result is less than the second threshold Re min , then a prompt to rebuild the AD / MCI large language model is returned (at this time, the clue information also includes the evaluation and scoring of the previous second interaction result). If the evaluation and scoring of the two (configurable) second interaction results are both less than the second threshold Re min , then return to step 1 to re-preprocess the question (at this time, the clue information also includes the evaluation and scoring of the previous second interaction results). If the evaluation and scoring of the second interaction results returned three times (configurable) are all less than the second threshold Re min , it will directly return to the user "cannot answer".

[0121] Furthermore, when performing consistency evaluation and scoring on the second interaction result, the following methods may be used, including but not limited to:

[0122] First, a key summary prompt is constructed based on the second interaction result, and the key summary prompt is input into the AD / MCI large language model to obtain a key summary result of the second interaction result.

[0123] In the specific implementation, according to the second interaction result Re for the question Query input by the user obtained in the above steps, a key summary prompt is constructed (including: role, task-requirement combined with user question, reference materials, second interaction result, etc. to generate a key summary), and the key summary prompt is sent to the AD / MCI large language model to generate a key summary result Re for the second interaction result. sum .

[0124] Then, the key point summary result is scored based on the pre-constructed third evaluation scoring model to obtain an evaluation score of the key point summary result.

[0125] In the specific implementation, the key points of the model feedback are summarized and the results are summarized. sumThe consistency evaluation score can be completed by the third evaluation scoring model built in advance. The evaluation score results are composed of two parts. On the one hand, it is given by the AD / MCI large language model that generates the key summary results. By constructing a prompt consisting of roles, tasks-providing a relevance score between 0 and 100, user questions, second interaction results and key summary results, and sending it to the AD / MCI large language model, the consistency score of the key summary results is obtained, thereby obtaining the relevance score for the key summary results and the second interaction results. model-sum On the other hand, the third evaluation scoring model sub-model that has been trained in advance (which can be tuned based on a pre-trained model such as T5) obtains the score of the relevance between the second interaction result and the key summary result. s-sum ; Finally, we get the key points of the AD / MCI large language model feedback and the evaluation score Relevance sum =(Relevance s-sum +Relevance model-sum ) / 2.

[0126] Finally, if the evaluation score of the key point summary result is greater than or equal to the third threshold, the second interaction result is fed back to the user; if the evaluation score of the key point summary result is less than the third threshold, the prompt of the AD / MCI large language model is reconstructed to obtain a new second interaction result. If the relevance evaluation score of the new second interaction result is greater than or equal to the second threshold and less than the first threshold, the new second interaction result is summarized and the key point summary result of the new second interaction result is scored. If the evaluation score of the key point summary result of the new second interaction result is less than the third threshold, the question is preprocessed again, and the preprocessed question is searched in the AD / MCI knowledge base and interacted with the AD / MCI large language model to obtain a new second interaction result.

[0127] In the specific implementation, the evaluation scores of the key points summary results are finally obtained and the situation is processed according to the third threshold value (Re sum-min ), if the evaluation score of the key summary result is greater than or equal to the third threshold Re sum -min , the second interaction result is directly fed back to the user; otherwise, a prompt to rebuild the AD / MCI large language model is returned (at this time, the clue information also includes the consistency evaluation score for the previous second interaction result). If the consistency evaluation scores of the key summary results of the second interaction results returned twice (configurable) are both less than the third threshold Re sum-min, then return to step 1 (at this time, the clue information also includes the consistency evaluation score for the previous second interaction result), if the consistency evaluation scores of the key summary results of the second interaction results returned three times (configurable) are all less than the third threshold Re sum -min , it will directly return to the user "cannot answer".

[0128] Further, for the second interaction result Re that has passed the evaluation and scoring, a reference mark is generated in the second interaction result according to each sentence in the recorded second interaction result and the corresponding reference material number, and the reference material content and reference material number are listed after the second interaction result Re, thereby generating the final second interaction result Re last , to provide feedback to users.

[0129] The above AD and MCI knowledge discovery method provided by the embodiment of the present invention has the following advantages and positive effects compared with the prior art:

[0130] (1) By constructing a real-time updated, multi-dimensional AD / MCI knowledge base including documents, entities, evidence, indexes, etc., the AD / MCI knowledge is ensured to be multi-dimensional, cross-granular, comprehensive and up-to-date; (2) Based on the knowledge structure of document tags, entities, evidence, etc. of the multi-dimensional knowledge base, and a fine-tuned large language model, a combination of knowledge and data-driven methods is adopted to ensure the automation and human-like effect of the AD / MCI knowledge discovery process; (3) In order to ensure that knowledge discovery is as accurate as possible, methods such as pre-judgment of user question queries (including answer method judgment) are adopted to match the complexity of questions with the answer method, thereby ensuring accuracy and effect while ensuring the efficiency of knowledge discovery; (4) By optimizing user question queries, optimizing knowledge retrieval, and optimizing prompt construction and interaction results, the accuracy of reply feedback results is further guaranteed; (5) By adopting a combination of user question queries and retrieval results, user question queries and other methods, the user query and search results are optimized. The method of evaluating and scoring the correlation between the user query and the interaction results, the retrieval results and the interaction results, and the interaction results and the key summary results ensures the self-consistency from the user query to the interaction results and the key summary results, including the method of giving an "unanswerable" interaction result in the case of a lower score, thereby ensuring the accuracy of the response interaction results at multiple levels; (6) Through step-by-step (loop-based) planning (based on the decomposition and processing of the problem, etc.), optimization (based on the evaluation of previous results), evaluation of interaction results and evaluation-based routing selection methods, the detailed decomposition and self-optimization of the knowledge discovery process are promoted, while ensuring the efficiency of the process, the quality and effect of the knowledge discovery process are also guaranteed; (7) By adding reference marks and reference materials to the content of the interaction results, the interpretability of the interaction results is enhanced, and the basis of the interaction results is explained; (8) While improving the efficiency, quality and effect of knowledge discovery by researchers in the AD / MCI field, the workload and complexity are also reduced.

[0131] For the AD and MCI knowledge discovery method provided in the above embodiment, the present invention also provides an AD and MCI knowledge discovery device, see Figure 2 The schematic diagram of the structure of an AD and MCI knowledge discovery device shown in FIG. 1 shows that the device mainly includes the following parts:

[0132] The knowledge base construction module 201 is used to obtain documents in the field of AD / MCI and construct an AD / MCI knowledge base based on the documents; wherein the AD / MCI knowledge base includes: an AD / MCI document base, an AD / MCI document entity base, an AD / MCI evidence base, and an AD / MCI index base;

[0133] The pre-judgment module 202 is used to obtain the question input by the user and pre-judgment the question to determine the way to answer the question;

[0134] The retrieval and interaction module 203 is used to perform AD / MCI knowledge base retrieval and / or AD / MCI large language model interaction based on the question answering method to obtain the question answering result; wherein the question answering result includes one of the following: a first final retrieval result, a first interaction result, a second final retrieval result and a second interaction result based on the second final retrieval result.

[0135] The result evaluation and feedback module 204 is used to feed back the first final retrieval result or the first interaction result to the user if a first final retrieval result or a first interaction result is obtained, and to perform relevance evaluation and consistency evaluation on the second interaction result if a second interaction result is obtained, and to feed back the second interaction result to the user if the score of the second interaction result meets the preset conditions.

[0136] The above-mentioned AD and MCI knowledge discovery device provided by the present invention ensures that AD / MCI knowledge is comprehensive and up-to-date by constructing a real-time updated, multi-dimensional AD / MCI knowledge base including documents, entities, evidence, indexes, etc.; by pre-judging the user's questions and matching the complexity of the questions with the answering methods, it ensures the accuracy and effectiveness while ensuring the efficiency of knowledge discovery; by performing relevance evaluation and consistency evaluation on the retrieval results of the user's questions and the interaction results of the AD / MCI large language model, it ensures the accuracy of the reply results, thereby improving the efficiency and quality of acquiring and discovering AD / MCI related knowledge and reducing the workload and complexity.

[0137] It should be noted that the implementation principle and technical effects of the device provided in the embodiment of the present invention are the same as those of the aforementioned method embodiment. For the sake of brief description, for matters not mentioned in the device embodiment, reference may be made to the corresponding contents in the aforementioned method embodiment.

[0138] An embodiment of the present invention further provides an electronic device. Specifically, the electronic device includes a processor and a storage device. The storage device stores a computer program, and when the computer program is executed by the processor, it executes the method described in any one of the above implementation methods.

[0139] Figure 3 A structural diagram of an electronic device provided in an embodiment of the present invention, the electronic device 100 includes: a processor 30, a memory 31, a bus 32 and a communication interface 33, wherein the processor 30, the communication interface 33 and the memory 31 are connected via the bus 32; the processor 30 is used to execute an executable module stored in the memory 31, such as a computer program.

[0140] The memory 31 may include a high-speed random access memory (RAM), and may also include a non-volatile memory, such as at least one disk memory. The communication connection between the system network element and at least one other network element is realized through at least one communication interface 33 (which may be wired or wireless), and the Internet, wide area network, local area network, metropolitan area network, etc. may be used.

[0141] The bus 32 may be an ISA bus, a PCI bus, or an EISA bus, etc. The bus may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 3 Only one bidirectional arrow is used in the diagram, but this does not mean that there is only one bus or only one type of bus.

[0142] Among them, the memory 31 is used to store programs, and the processor 30 executes the program after receiving the execution instruction. The method executed by the device for flow process definition disclosed in any embodiment of the above-mentioned embodiments of the present invention can be applied to the processor 30 or implemented by the processor 30.

[0143] The processor 30 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the hardware integrated logic circuit or software instructions in the processor 30. The above processor 30 can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The methods, steps and logic block diagrams disclosed in the embodiments of the present invention can be implemented or executed. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor. The steps of the method disclosed in the embodiments of the present invention can be directly embodied as a hardware decoding processor to be executed, or the hardware and software modules in the decoding processor can be executed. The software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory 31, and the processor 30 reads the information in the memory 31 and completes the steps of the above method in combination with its hardware.

[0144] The computer program product of the readable storage medium provided in the embodiment of the present invention includes a computer-readable storage medium storing program code, and the instructions included in the program code can be used to execute the methods described in the previous method embodiments. The specific implementation can be referred to the previous method embodiments, which will not be repeated here.

[0145] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program codes.

[0146] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-mentioned embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-mentioned embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A method for discovering AD and MCI knowledge, characterized in that: include: Acquire literature in the field of AD / MCI, and build an AD / MCI knowledge base based on the literature; wherein the AD / MCI knowledge base includes: an AD / MCI literature base, an AD / MCI literature entity base, an AD / MCI evidence base, and an AD / MCI index base; Obtaining questions input by users, and pre-judging the questions to determine how to answer them; Based on the question answering method, the AD / MCI knowledge base search and / or AD / MCI large language model interaction are performed to obtain a question answering result; wherein the question answering result includes one of the following: a first final search result, a first interaction result, a second final search result, and a second interaction result based on the second final search result; If the first final retrieval result or the first interaction result is obtained, the first final retrieval result or the first interaction result is fed back to the user; if the second interaction result is obtained, the second interaction result is scored for relevance and consistency; if the score of the second interaction result meets the preset conditions, the second interaction result is fed back to the user.

2. The method according to claim 1, characterized in that Based on the literature, an AD / MCI knowledge base is constructed, including: According to the labels of the knowledge system of the pre-constructed AD / MCI document library, the metadata of each document is extracted, and the metadata is processed based on the pre-constructed knowledge rules, classification models and information extraction models to obtain the content corresponding to each label in the knowledge system, and the AD / MCI document library is constructed according to the content corresponding to each label of each document and the metadata; wherein the metadata also includes the document structure and content; For each document in the AD / MCI document library, entities in each document are extracted based on a pre-built entity information extraction model to build an AD / MCI document entity library; For each document in the AD / MCI document library, if the document type tag content of the document includes RCT, then extract the evidence information in each document based on the pre-built evidence extraction model to build the AD / MCI evidence library; wherein the evidence information is information related to RCT; Sub-indexes of the AD / MCI document library, the AD / MCI document entity library and the AD / MCI evidence library are respectively constructed, and the AD / MCI document library, the AD / MCI document entity library and the AD / MCI evidence library are integrated and associated based on the association IDs between the data of the AD / MCI document library, the AD / MCI document entity library and the AD / MCI evidence library to construct an overall index to obtain the AD / MCI index library; wherein the sub-index and the overall index include an inverted index and a vector index.

3. The method according to claim 1, characterized in that The method of answering the question includes: an AD / MCI knowledge base retrieval method, an AD / MCI large language model interactive answer method, and a combination of an AD / MCI knowledge base retrieval and an AD / MCI large language model interactive answer method; Based on the question answering method, the AD / MCI knowledge base search and / or AD / MCI large language model interaction are performed to obtain the question answering result, including: If the method of answering the question is a method of searching the AD / MCI knowledge base, searching the AD / MCI knowledge base based on the question to obtain a first final search result; If the way of answering the question is the way of interactive answering by the AD / MCI large language model, inputting the question into the AD / MCI large language model to obtain a first interactive result; If the method of answering the question is a combination of AD / MCI knowledge base retrieval and AD / MCI large language model interactive answering, the question is preprocessed, and the AD / MCI knowledge base is searched based on the preprocessed question to obtain a second final retrieval result, and a prompt of the AD / MCI large language model is constructed based on the second final retrieval result, and the prompt is input into the AD / MCI large language model to obtain a second interactive result based on the second final retrieval result.

4. The method according to claim 3, characterized in that The question is preprocessed, and an AD / MCI knowledge base search is performed based on the preprocessed question to obtain a second final search result, including: Inputting the question into the AD / MCI large language model to decompose the question, and concatenating the decomposition result with the question to obtain a concatenated query question; Performing word segmentation and keyword extraction on the concatenated query question to obtain a first keyword list, and inputting the first keyword list into a pre-built synonym large language model to obtain a first synonym list of keywords; Extracting entities of the question based on a pre-built information extraction model to obtain a first entity list; Performing search condition judgment on the question based on a pre-built judgment model to obtain a judgment condition result; wherein the search condition at least includes: the document type to which the target content belongs and the location of the target content in the document; Determine a first search logic for an inverted index of the AD / MCI document library based on the first synonym list of the keyword and the judgment condition result, and determine a second search logic for a vector index of the AD / MCI document library based on the concatenated query question and the judgment condition result, perform an inverted index search based on the first search logic to obtain a first document search result, and perform a vector search based on the second search logic to obtain a second document search result; Determine a third search logic for the AD / MCI document entity library based on the first entity list and the judgment condition result, and perform an AD / MCI document entity library search based on the third search logic to obtain a first entity search result; If the document type in the judgment condition result includes RCT, determining a fourth search logic for the AD / MCI evidence base based on the first entity list and the first synonym list of the keyword, and searching the AD / MCI evidence base based on the fourth search logic to obtain a first evidence search result; The question, the first document search result, the second document search result, the first entity search result and the first evidence search result are input into a pre-constructed first evaluation scoring model to obtain a scoring result for each search result, and a second final search result is determined based on the scoring result.

5. The method according to claim 4, characterized in that The prompt for constructing the AD / MCI large language model based on the second final search result includes: Determine the roles, tasks and requirements of the AD / MCI large language model; After preprocessing the second final search result, the preprocessed second final search result is sorted based on the scoring result of the second final search result to obtain reference materials; wherein the preprocessing of the second final search result includes: summarizing and compressing the document content in the second final search result, compressing the content in the second final search result that exceeds a preset length, and combining the evidence content in the second final search result into a natural language sentence; Determining clue information based on the first keyword list and the first entity list; A prompt for the AD / MCI large language model is constructed based on the question, the role, task, requirement, reference material and clue information of the AD / MCI large language model.

6. The method according to claim 1, characterized in that The second interaction result is scored for relevance evaluation and consistency evaluation, and if the score of the second interaction result meets a preset condition, the second interaction result is fed back to the user, including: filtering the second interaction result based on a preset rule; Inputting the filtered second interaction result into a pre-constructed second evaluation and scoring model to obtain an evaluation and scoring of the second interaction result; If the evaluation score of the second interaction result is greater than or equal to the first threshold, feeding back the second interaction result to the user; If the evaluation score of the second interaction result is greater than or equal to a second threshold and less than the first threshold, then the second interaction result is evaluated and scored for consistency; wherein the first threshold is greater than the second threshold; If the evaluation score of the second interaction result is less than the second threshold, the prompt of the AD / MCI large language model is reconstructed to obtain a new second interaction result, and the new second interaction result is evaluated and scored for relevance; if the evaluation score of the new second interaction result is less than the second threshold, the question is preprocessed again, and the preprocessed question is searched through the AD / MCI knowledge base and interacted with the AD / MCI large language model to obtain a new second interaction result.

7. The method according to claim 6, characterized in that The consistency evaluation and scoring of the second interaction result includes: constructing a key summary prompt based on the second interaction result, and inputting the key summary prompt into the AD / MCI large language model to obtain a key summary result of the second interaction result; Scoring the key point summary result based on a pre-constructed third evaluation scoring model to obtain an evaluation score of the key point summary result; If the evaluation score of the key point summary result is greater than or equal to a third threshold, feeding back the second interaction result to the user; If the evaluation score of the key point summary result is less than the third threshold, the prompt of the AD / MCI large language model is reconstructed to obtain a new second interaction result, and the key points of the new second interaction result are summarized and the key point summary result of the new second interaction result is scored. If the evaluation score of the key point summary result of the new second interaction result is less than the third threshold, the question is preprocessed again, and the preprocessed question is searched in the AD / MCI knowledge base and interacted with the AD / MCI large language model to obtain a new second interaction result.

8. An AD and MCI knowledge discovery device, characterized in that: include: A knowledge base construction module is used to obtain literature in the field of AD / MCI and construct an AD / MCI knowledge base based on the literature; wherein the AD / MCI knowledge base includes: an AD / MCI literature base, an AD / MCI literature entity base, an AD / MCI evidence base, and an AD / MCI index base; A pre-judgment module is used to obtain questions input by users and pre-judgment the questions to determine the way to answer the questions; A retrieval and interaction module, configured to perform the AD / MCI knowledge base retrieval and / or AD / MCI large language model interaction based on the question answering method to obtain a question answering result; wherein the question answering result includes one of the following: a first final retrieval result, a first interaction result, a second final retrieval result, and a second interaction result based on the second final retrieval result; A result evaluation and feedback module is used to feed back the first final retrieval result or the first interaction result to the user if the first final retrieval result or the first interaction result is obtained, and to perform relevance evaluation and consistency evaluation on the second interaction result if the second interaction result is obtained, and to feed back the second interaction result to the user if the score of the second interaction result meets a preset condition.

9. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory stores computer executable instructions that can be executed by the processor, and the processor executes the computer executable instructions to implement the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are performed.

Images