A medicine knowledge question and answer method, device, storage medium and equipment

Abstract

This application discloses a drug knowledge question-answering method, apparatus, storage medium, and device. The method includes: firstly, decomposing the target drug question text to obtain N target sub-question texts; extracting pharmaceutical entities from them; then concatenating the two in pairs to obtain each pharmaceutical entity and target sub-question pair, and classifying their intents to obtain a first intent classification result for each pharmaceutical entity; next, using preset pharmaceutical rules, classifying the intents of the target sub-questions in each pharmaceutical entity and target sub-question pair to obtain a second intent classification result for each pharmaceutical entity; merging this second and first intent classification results to obtain a final intent classification result for each pharmaceutical entity; then, using the drug instructions and / or a preset drug knowledge base for knowledge retrieval, and inputting the retrieved knowledge, combined with prompts, into a preset large language model to obtain a more accurate response to the target drug question text output by the model.

Description

Technical Field

[0001] This application relates to the field of natural language processing technology, and in particular to a method, apparatus, storage medium and device for drug knowledge question answering. Background Technology

[0002] With the rapid development of information technologies such as artificial intelligence and the Internet of Things, the application scenarios of human-computer interaction are becoming increasingly widespread. Various intelligent interactive software and devices are appearing in people's lives and work, such as Chat Generative Pre-trained Transformer (ChatGPT), smart speakers, and smart TVs. These devices can provide intelligent interactive functions for many application scenarios, such as information retrieval and knowledge Q&A, to assist users in completing various behavioral intentions.

[0003] Currently, when users inquire about drug-related questions via smart interactive software or devices (such as smart speakers and smart TVs), they often need to provide corresponding evidence of medication use; otherwise, the responses are unlikely to be convincing. Therefore, it is necessary to introduce rational drug use plugins to provide key knowledge information and reference evidence for conclusions in responses to drug-related questions. However, existing large-scale drug knowledge question-answering plugins typically process knowledge content rather crudely by slicing and retrieving knowledge documents. This may result in incomplete knowledge fragments or under-segmentation of knowledge, compromising the quality of the knowledge content and reducing the accuracy of responses to drug-related questions, thereby degrading the user's interactive experience. Summary of the Invention

[0004] The main objective of this application is to provide a drug knowledge question-and-answer method, apparatus, storage medium, and device that can improve the accuracy of responses to drug-related questions input by users into intelligent interactive software or devices, thereby enhancing the user's interactive experience.

[0005] This application provides a multi-drug knowledge question-answering method, including:

[0006] Obtain the target drug question text to be answered; decompose the target drug question text into N target sub-question texts; and extract pharmaceutical entities from the N target sub-question texts to obtain the pharmaceutical entities in the N target sub-question texts; where N is a positive integer greater than 0;

[0007] The pharmaceutical entity and the N target sub-problems are concatenated in pairs to obtain each pharmaceutical entity and target sub-problem pair. The intent of each pharmaceutical entity and target sub-problem pair is then classified to obtain the first intent classification result of each pharmaceutical entity.

[0008] Using preset pharmaceutical rules, the intention classification of the target sub-problems in each pharmaceutical entity and target sub-problem pair is performed to obtain the second intention classification result of each pharmaceutical entity; and then the second intention classification result is merged with the first intention classification result to obtain the intention classification result of each pharmaceutical entity.

[0009] Based on the intent classification results of each pharmaceutical entity, knowledge retrieval is performed using the drug instructions and / or a preset drug knowledge base. The retrieved knowledge, combined with the prompt instruction, is then input into a preset large language model to obtain the response content of the large language model to the target drug question text.

[0010] In one possible implementation, the decomposition of the target drug problem text into N target sub-problem texts includes:

[0011] Combining the sub-problem splitting prompt instruction, the target drug problem text is decomposed using a preset large language model to obtain N target sub-problem texts;

[0012] The large language model is obtained by training language rules and patterns using a large-scale language dataset through autoregressive generation.

[0013] In one possible implementation, the pharmaceutical entity includes a drug entity, a disease entity, and a symptom entity.

[0014] In one possible implementation, the step of using preset pharmaceutical rules to classify the intent of the target sub-problems in each pair of pharmaceutical entities and target sub-problems to obtain a second intent classification result for each pharmaceutical entity includes:

[0015] Patient information entities are extracted from the N target sub-problem texts to obtain patient information entities in the N target sub-problem texts; and patient profiles are constructed using the patient information entities and pharmaceutical entities;

[0016] By using reasoning logic pre-constructed based on doctors' experience, combined with a drug knowledge base, patient profiles are analyzed and compared to entities to arrive at conclusions on whether each drug should be used.

[0017] Based on the conclusions regarding whether each drug should be used, the content for knowledge recommendation is determined for the corresponding drug, and the recommended knowledge content is used as the second intent classification result for the corresponding drug entity.

[0018] In one possible implementation, the step of performing knowledge retrieval based on the intent classification results of each of the pharmaceutical entities, using drug instructions and / or a pre-defined drug knowledge base, includes:

[0019] Using drug instructions and / or a pre-set drug knowledge base, entity alignment and disambiguation are performed on each pharmaceutical entity to obtain the corresponding generic drug name, standard disease term, and standard symptom term.

[0020] Based on the intent classification results of each pharmaceutical entity, the unstructured, semi-structured, and structured knowledge in the drug instructions and / or the preset drug knowledge base are queried as the search results.

[0021] In one possible implementation, the step of querying unstructured, semi-structured, and structured knowledge in the drug instructions and / or a pre-defined drug knowledge base based on the intent classification results of each of the pharmaceutical entities, as the retrieval results, includes:

[0022] Based on the intent classification results of drug entities, query unstructured knowledge in the drug instructions as the search results;

[0023] And / or, based on the intent classification results of disease entities, query the structured knowledge in the preset drug knowledge base as the search results;

[0024] And / or, based on the intent classification results of each of the pharmaceutical entities, query unstructured knowledge in a pre-defined drug knowledge base using sparse representation as the retrieval result.

[0025] In one possible implementation, when knowledge retrieval is performed using the drug instructions and / or a pre-defined drug knowledge base based on the intent classification results of each pharmaceutical entity, and no knowledge is found, the method further includes:

[0026] Knowledge is searched using a pre-set web search engine, and the knowledge content in the search results is located using the large language model.

[0027] This application also provides a drug knowledge question-and-answer device, including:

[0028] An acquisition unit is configured to acquire the target drug question text to be answered; decompose the target drug question text into N target sub-question texts; and extract pharmaceutical entities from the N target sub-question texts to obtain the pharmaceutical entities in the N target sub-question texts; wherein N is a positive integer greater than 0.

[0029] The first classification unit is used to concatenate the pharmaceutical entity and the N target sub-problems in pairs to obtain each pharmaceutical entity and target sub-problem pair, and to perform intent classification on each pharmaceutical entity and target sub-problem pair to obtain the first intent classification result of each pharmaceutical entity;

[0030] The second classification unit is used to classify the intentions of the target sub-problems in each pharmaceutical entity and target sub-problem pair using preset pharmaceutical rules, to obtain the second intention classification result of each pharmaceutical entity; and to merge it with the first intention classification result to obtain the intention classification result of each pharmaceutical entity.

[0031] The retrieval unit is used to perform knowledge retrieval based on the intent classification results of each pharmaceutical entity, using the drug instructions and / or a preset drug knowledge base, and inputting the retrieved knowledge, combined with the prompt instruction, into a preset large language model to obtain the response content of the large language model outputting the question text for the target drug.

[0032] In one possible implementation, the acquisition unit is specifically used for:

[0033] Combining the sub-problem splitting prompt instruction, the target drug problem text is decomposed using a preset large language model to obtain N target sub-problem texts;

[0034] The large language model is obtained by training language rules and patterns using a large-scale language dataset through autoregressive generation.

[0035] In one possible implementation, the pharmaceutical entity includes a drug entity, a disease entity, and a symptom entity.

[0036] In one possible implementation, the second classification unit includes:

[0037] An extraction subunit is used to extract patient information entities from the N target sub-question texts to obtain patient information entities in the N target sub-question texts; and to construct a patient profile using the patient information entities and pharmaceutical entities;

[0038] The comparison subunit is used to analyze and compare patient profiles using reasoning logic pre-built based on doctors' experience, combined with a drug knowledge base, to arrive at a conclusion on whether each drug should be used.

[0039] The determination subunit is used to determine the knowledge recommendation content for the corresponding drug based on the conclusion of whether each drug should be used, and to use the recommended knowledge content as the second intent classification result of the corresponding drug entity.

[0040] In one possible implementation, the retrieval unit includes:

[0041] Obtain sub-units to use drug instructions and / or a preset drug knowledge base to perform entity alignment and entity disambiguation on each pharmaceutical entity to obtain the corresponding generic drug name, standard disease term, and standard symptom term;

[0042] The query subunit is used to query unstructured, semi-structured, and structured knowledge in drug instructions and / or a preset drug knowledge base based on the intent classification results of each of the pharmaceutical entities, and use this as the retrieval result.

[0043] In one possible implementation, the query subunit is specifically used for:

[0044] Based on the intent classification results of drug entities, query unstructured knowledge in the drug instructions as the search results;

[0045] And / or, based on the intent classification results of disease entities, query the structured knowledge in the preset drug knowledge base as the search results;

[0046] And / or, based on the intent classification results of each of the pharmaceutical entities, query unstructured knowledge in a pre-defined drug knowledge base using sparse representation as the retrieval result.

[0047] In one possible implementation, the device further includes:

[0048] The search unit is used to perform knowledge retrieval using drug instructions and / or a preset drug knowledge base based on the intent classification results of each pharmaceutical entity. If no knowledge is found, it performs knowledge search through a preset web search engine and uses the large language model to locate the knowledge content in the search results.

[0049] This application also provides a drug knowledge question-and-answer device, including: a processor, a memory, and a system bus;

[0050] The processor and the memory are connected via the system bus;

[0051] The memory is used to store one or more programs, the one or more programs including instructions, which, when executed by the processor, cause the processor to perform any of the above-described implementations of the drug knowledge question-and-answer method.

[0052] This application also provides a computer-readable storage medium storing instructions that, when executed on a terminal device, cause the terminal device to perform any of the above-described drug knowledge question-and-answer methods.

[0053] This application also provides a computer program product, which, when run on a terminal device, causes the terminal device to execute any of the above-described drug knowledge question-and-answer methods.

[0054] This application provides a drug knowledge question-answering method, apparatus, storage medium, and device. First, it acquires the target drug question text to be answered; then, it decomposes the target drug question text into N target sub-question texts; and extracts pharmaceutical entities from the N target sub-question texts to obtain the pharmaceutical entities within the N target sub-question texts; then, it concatenates the pharmaceutical entities and the N target sub-questions pair by pair to obtain each pharmaceutical entity and target sub-question pair, and performs intent classification on each pharmaceutical entity and target sub-question pair to obtain a first intent classification result for each pharmaceutical entity; next, it uses preset pharmaceutical rules to perform intent classification on the target sub-questions in each pharmaceutical entity and target sub-question pair to obtain a second intent classification result for each pharmaceutical entity; and merges this second and first intent classification results to obtain the intent classification result for each pharmaceutical entity; then, based on the intent classification results of each pharmaceutical entity, it performs knowledge retrieval using the drug instructions and / or a preset drug knowledge base, and inputs the retrieved knowledge, combined with a prompt instruction, into a preset large language model to obtain the response content for the target drug question text output by the large language model.

[0055] As can be seen, since this application first decomposes the target drug question text into sub-questions and extracts pharmaceutical entities, it is beneficial to achieve fine-grained classification of the target user's question intent. Then, it integrates the drug knowledge question-answering plugin with the preset pharmaceutical rules and the preset large language model. Based on the intent classification results of each pharmaceutical entity, it uses the drug instructions and / or the preset drug knowledge base to perform more accurate knowledge retrieval, thereby effectively improving the accuracy of subsequent answers to the target drug question and thus improving the interactive experience of the target user. Attached Figure Description

[0056] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0057] Figure 1 A flowchart illustrating a drug knowledge question-and-answer method provided in an embodiment of this application;

[0058] Figure 2 This is an example diagram illustrating how, based on the conclusion that each drug should be used, knowledge recommendations are made for the corresponding drug, and the recommended knowledge content is used as the second intent classification result of the corresponding drug entity, as provided in the embodiments of this application.

[0059] Figure 3An example diagram illustrating the use of drug instructions and / or a pre-defined drug knowledge base for knowledge retrieval based on the intent classification results of each pharmaceutical entity, as provided in the embodiments of this application.

[0060] Figure 4 This is a schematic diagram of the composition of a drug knowledge question-and-answer device provided in an embodiment of this application. Detailed Implementation

[0061] As is well known, drug-related questions are common and key issues in medical Q&A scenarios. Currently available intelligent interactive software or devices, such as large language models (LLMs), cannot accurately answer questions about actual drug interactions, contraindications, and dosage. Furthermore, responses to drug-related questions need to provide corresponding evidence of medication use; otherwise, the responses are unlikely to be convincing.

[0062] Therefore, a rational drug use plugin is needed to provide key knowledge information for answering drug-related questions and to offer reference for conclusions. However, existing intelligent interactive software or devices using large language models typically employ a slice-based retrieval method for knowledge documents. This involves dividing text containing the required knowledge into shorter natural language segments, filtering out the most relevant segments based on text similarity or other rules during question answering, and then inputting these segments into the large model so that the model can effectively utilize this knowledge in the question-answering process. However, this approach presents three main problems:

[0063] First, using text segmentation to process knowledge content is rather crude, and may result in incomplete knowledge fragments or under-segmentation of knowledge, which makes it impossible to guarantee the quality of the knowledge content.

[0064] Secondly, knowledge fragments selected by simply using methods such as text similarity comparison or keyword matching usually have poor accuracy and recall, and the problem of knowledge mis-absorption is quite serious, making it impossible to accurately obtain the knowledge fragments that are truly needed.

[0065] Third, the knowledge queried lacks structure and is difficult to trace. Regarding the risks of drug use, it is difficult to determine the knowledge basis of intelligent interactive software or devices such as LLM when they answer, making their answers lack credibility and persuasiveness.

[0066] In addition, current rational drug use management or inference systems utilize pharmaceutical professionals to transform their accumulated pharmaceutical knowledge into knowledge rules. Artificial intelligence or traditional software technology then constructs a drug knowledge base and answers questions regarding dosage, contraindications, and interactions in medication regimens. However, because these rational drug use management inference systems lack question-and-answer capabilities, they are typically only used in internal professional settings such as hospitals. They lack the ability to connect with a broad range of general users and answer drug-related questions, as offered by intelligent interactive software or device plugins like LLM (Limited Management System).

[0067] To address the aforementioned deficiencies, this application provides a drug knowledge question-answering method. First, the target drug question text to be answered is obtained; then, the target drug question text is decomposed into N target sub-question texts; pharmaceutical entities are extracted from the N target sub-question texts to obtain the pharmaceutical entities within the N target sub-question texts; then, the pharmaceutical entities and the N target sub-questions are concatenated pairwise to obtain each pharmaceutical entity-target sub-question pair, and intent classification is performed on each pharmaceutical entity-target sub-question pair to obtain a first intent classification result for each pharmaceutical entity; next, using preset pharmaceutical rules, intent classification is performed on the target sub-questions in each pharmaceutical entity-target sub-question pair to obtain a second intent classification result for each pharmaceutical entity; this second intent classification result is then merged with the first intent classification result to obtain the intent classification result for each pharmaceutical entity; finally, based on the intent classification results of each pharmaceutical entity, knowledge retrieval is performed using the drug instructions and / or a preset drug knowledge base, and the retrieved knowledge, combined with a prompt instruction, is input into a preset large language model to obtain the response content for the target drug question text output by the large language model.

[0068] As can be seen, since this application first decomposes the target drug question text into sub-questions and extracts pharmaceutical entities, it is beneficial to achieve fine-grained classification of the target user's question intent. Then, it integrates the drug knowledge question-answering plugin with the preset pharmaceutical rules and the preset large language model. Based on the intent classification results of each pharmaceutical entity, it uses the drug instructions and / or the preset drug knowledge base to perform more accurate knowledge retrieval, thereby effectively improving the accuracy of subsequent answers to the target drug question and thus improving the interactive experience of the target user.

[0069] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0070] First Embodiment

[0071] See Figure 1 This is a flowchart illustrating a drug knowledge question-and-answer method provided in this embodiment. The method includes the following steps:

[0072] S101: Obtain the target drug question text to be answered; decompose the target drug question text to obtain N target sub-question texts; and extract pharmaceutical entities from the N target sub-question texts to obtain the pharmaceutical entities in the N target sub-question texts; where N is a positive integer greater than 0.

[0073] In this embodiment, any drug-related question text that requires a response and is input by the user into intelligent interactive software or devices such as LLM is defined as the target drug-related question text to be responded to. Furthermore, this embodiment does not limit the language type of the target drug-related question text; for example, it can be Chinese or English text. This embodiment also does not limit the content of the target drug-related question text; that is, the target drug-related question text can be any drug-related question text in the medical field.

[0074] Understandably, the target drug question text can be a sentence, a collection of words. After obtaining the target drug question text to be answered, a sub-question splitting prompt can be used, along with a pre-defined large language model, to decompose the target drug question text into N (N is a positive integer greater than 0) target sub-question texts with relatively clearer and more singular intents. If the target drug question text cannot be split, then that target drug question text can be used as the unique target sub-question text.

[0075] In this application, the specific content of the sub-problem splitting prompt instruction is not limited. It can be set according to the actual situation and experience. For example, the sub-problem splitting prompt instruction can be, but is not limited to, "Split the user's question into multiple sub-problems and give an importance score for each sub-problem. The user's question is: [Question]", where "[Question]" is a slot to be filled with the text of the target drug question to be answered.

[0076] Large Language Models (LLMs) are deep learning-based language models that can generate new language expressions, such as text, sentences, paragraphs, and even articles, based on input text content. LLMs utilize large-scale language datasets and are trained on language rules and patterns through autoregressive generation. They can simulate human commands to generate language expressions (such as text data). Specifically, when generating new text data, LLMs predict the probability of the next language unit based on previously generated content until complete text data is generated.

[0077] During model training, a sample Chinese text can be extracted from the training data to construct a sample prompt template text, which is then used as the model input. The model's output translation prediction is compared with the manually labeled translation results corresponding to the recorded sample Chinese text, and the model parameters are updated based on the differences. This process is repeated multiple times until preset conditions are met (such as reaching a preset number of training iterations or the optimization function value remaining essentially unchanged). At this point, updating the model parameters stops, completing the training of the Large Language Model (LLM) and generating a trained LLM.

[0078] Simultaneously, existing or future Named Entity Recognition (NER) methods can be used to extract pharmaceutical entities from the N target sub-problem texts, obtaining the pharmaceutical entities in the N target sub-problem texts for subsequent step S102. These pharmaceutical entities include, but are not limited to, drug entities, disease entities, and symptom entities.

[0079] Specifically, the obtained target drug question text to be answered can be input into a small entity extraction model based on BERT (Bidirectional Encoder Representations for Transformers) modeling to extract pharmaceutical entities, resulting in four types of pharmaceutical entities in each target sub-question text: drug, drug type, disease, and symptom. Then, the drug type entity is converted into a most common specific drug entity for subsequent steps; for example, the drug type entity "anti-inflammatory drug" can be converted into the drug entity "acetaminophen tablets" for further processing.

[0080] In this way, by splitting the target drug question text into multiple target sub-question texts and classifying the query intent based on each medical entity in the sub-question, fine-grained analysis of user intent can be achieved. This lays the foundation for accurate knowledge matching of the large language model plugin in subsequent steps and solves the problem of unclear knowledge requirements of the plugin.

[0081] S102: Pairwise concatenate the pharmaceutical entity and N target sub-problems to obtain each pair of pharmaceutical entity and target sub-problems, and perform intent classification on each pair of pharmaceutical entity and target sub-problems to obtain the first intent classification result of each pharmaceutical entity.

[0082] In this embodiment, after obtaining N target sub-question texts and each pharmaceutical entity through step S101, in order to improve the accuracy of the response to the target drug question, each pharmaceutical entity and the N target sub-questions can be concatenated in pairs to obtain each pharmaceutical entity and target sub-question pair. The intent classification of each pharmaceutical entity and target sub-question pair can be performed. For example, each pharmaceutical entity and target sub-question pair can be input into a sentence intent classification mini-model built based on the BERT model to obtain the intent classification result of each pharmaceutical entity in the target sub-question sentence, and it is defined as the first intent classification result for executing the subsequent step S103.

[0083] For example, the pharmaceutical entity and the target sub-problem can be used to describe "pharmaceuticals". <sep>Reba Pat Film <sep>What is the dosage and usage of Rebamipide tablets? The input is used to classify the sentence intent using a small BERT-based intent classification model. Scores are obtained for each intent label, and the "dosage and usage" label with a score exceeding a preset threshold (e.g., 0.5) is taken as the intent label for the drug entity "Rebamipide tablets" in the sub-question "Orange" (i.e., the first intent classification result). It is understood that multiple intent labels can be taken for the same entity. <sep>The delimiter can be represented by other symbols, and this application does not limit the use of such symbols.

[0084] S103: Using preset pharmaceutical rules, classify the intentions of each pharmaceutical entity and target sub-problem pair to obtain the second intention classification result of each pharmaceutical entity; and merge it with the first intention classification result to obtain the intention classification result of each pharmaceutical entity.

[0085] It should be noted that, due to the limited logical reasoning capabilities of intelligent interactive software or devices such as large language models, for questions requiring the establishment of rules based on professional knowledge, such as the contraindications, indications, and interactions of drugs, it is difficult to directly provide affirmative answers based on knowledge alone. This results in vague and ambiguous responses from such software or devices. Therefore, this application proposes combining pre-set pharmaceutical rules with intelligent interactive software or devices such as large language models to alleviate the lack of pharmaceutical professional logical capabilities in these systems.

[0086] In this embodiment, preset pharmaceutical rules can be used to classify the intent of each pharmaceutical entity and target sub-problem pair, obtaining the intent classification result of each pharmaceutical entity (defined here as the second intent classification result). This result is then merged with the first intent classification result to obtain the final intent classification result for each pharmaceutical entity. Specifically, for labels such as contraindications, indications, and interactions of drugs that can be clearly concluded based on the questions, combined with the extracted pharmaceutical entity information, target sub-problem content, and physician experience rules, a conclusion inference is made. Recommendations and reasons are provided regarding whether to recommend the use of the mentioned drug, in order to execute subsequent step S104.

[0087] Specifically, one possible implementation is that the process of "using preset pharmaceutical rules to classify the intentions of each pharmaceutical entity and target sub-problem pair to obtain the second intention classification result of each pharmaceutical entity" in step S103 can include: first, extracting patient information (such as age, gender, etc.) entities from the N target sub-problem texts to obtain patient information entities in the N target sub-problem texts; and then using patient information entities and pharmaceutical entities (disease, symptom entities, etc.) to construct a patient profile.

[0088] Then, using reasoning logic pre-constructed based on doctors' experience, combined with a drug knowledge base, the patient profile is analyzed and compared, such as comparing drug and patient information, and comparing drugs with each other, to determine whether each drug should be used. For example, for the target sub-question text "Can I use Gastrointestinal Relief Capsules for constipation?", the query finds that "constipation" is among the contraindications for "Gastrointestinal Relief Capsules", so the conclusion is "there is a contraindication" and it should not be used.

[0089] Next, after formalizing the conclusions regarding whether each drug should be used—for example, obtaining the written conclusion, "It is not recommended to use Gastrointestinal Relief Capsules because the patient suffers from constipation, which is a contraindication to this drug"—this conclusion can be used to determine the content for knowledge recommendations for the corresponding drug, and the recommended knowledge content can be used as the second intent classification result for the corresponding drug entity. For example, when it is determined that a certain drug can be used, the corresponding drug's "Precautions" and "Adverse Reactions" can be recommended; when it is determined that there is a drug contraindication, the drug's "Contraindications" can be recommended. For example, the rule-based judgment process for "Indications" and "Contraindications" can be as follows: Figure 2 As shown.

[0090] S104: Based on the intent classification results of each pharmaceutical entity, perform knowledge retrieval using the drug instructions and / or a preset drug knowledge base, and input the retrieved knowledge, combined with the prompt instruction, into a preset large language model to obtain the response content of the target drug question text output by the large language model.

[0091] It should be noted that, in order to achieve 2. Highly effective knowledge matching, this application proposes to use semi-structured drug instructions, structured drug treatment recommendation data, etc. as the main knowledge sources, and to achieve highly effective knowledge matching based on the intent classification results of each pharmaceutical entity in the question. This reduces the misuse of knowledge caused by mismatch and omission of knowledge matching, thereby improving the accuracy of answers from large language models, etc.

[0092] Therefore, in this embodiment, after obtaining the intent classification results of each pharmaceutical entity through step S103, in order to improve the accuracy of the response to the target drug question, knowledge retrieval can be performed using the drug instructions and / or a preset drug knowledge base based on the intent classification results of each pharmaceutical entity. The retrieved knowledge is then combined with the prompt instruction and input into a preset large language model to obtain the response content for the target drug question text output by the large language model.

[0093] Specifically, firstly, the drug instructions and / or a pre-set drug knowledge base can be used to perform entity alignment and disambiguation on various pharmaceutical entities to obtain the corresponding generic drug name, standard disease terminology, and standard symptom terminology. Then, based on the intent classification results of each pharmaceutical entity, unstructured, semi-structured, and structured knowledge in the drug instructions and / or the pre-set drug knowledge base can be queried as search results. Subsequently, these results can be arranged and merged according to the text order of the corresponding target sub-questions. The merged knowledge list, combined with the corresponding question-answering prompt, is used as input to the large language model.

[0094] Specifically, based on the intent classification results of drug entities, unstructured knowledge in the drug instructions can be queried as the retrieval result; and / or, based on the intent classification results of disease entities, structured knowledge in a preset drug knowledge base can be queried as the retrieval result; and / or, based on the intent classification results of each of the pharmaceutical entities, unstructured knowledge in a preset drug knowledge base can be queried based on sparse representation as the retrieval result.

[0095] One possible implementation is, such as Figure 3 As shown, for drug entities not found in the pre-defined drug knowledge base, we can first try aligning them to existing drug entities in the knowledge base, calculating the number of misspellings, extra words, and missing words as 1, to correct erroneous descriptions of drug names in the target drug problem text. Simultaneously, we can map synonyms, brand names, and international nonproprietary names of drug entities to generic drug names using knowledge base resource data, and map disease entities and symptom entities to standard terms. For example, the drug entity "Claritin" is mapped to the generic drug name "Loratadine Tablets," and the disease entity "mild cerebral infarction" is mapped to the standard term "mild cerebral infarction."

[0096] Then, for drug entities, the corresponding items in the drug instructions can be queried based on intent tags as knowledge content. In particular, for semi-structured knowledge of usage and dosage, the drug's specifications can be added, such as "For compound donkey-hide gelatin granules with a specification of 4g / bag, the usage and dosage is: dissolve in boiling water and take 4 grams (1 bag) at a time, 3 times a day." For disease and symptom entities, the Top M items (the specific value is not limited, such as 3) of drug recommendations for the corresponding disease and symptom can be queried from the structured knowledge base and expressed using natural language, such as "Drugs that can be used for otitis media include amoxicillin (a penicillin antibiotic), cephalexin (a cephalosporin antibiotic), and levofloxacin hydrochloride ear drops (an otopathic antibiotic), etc." At the same time, the precautions for these Top M items of drugs can be added to the candidate knowledge list from the drug instructions.

[0097] For pharmaceutical entities for which there is no semi-structured or structured knowledge in the knowledge base, an inverted index query based on sparse representation can be performed from unstructured knowledge. Specifically, the query intent and keywords and extended terms from any knowledge document in the knowledge base are first identified and generated. The input content can consist of the pharmaceutical entity, intent tag, and target sub-question text; for example, the input content could be "amoxicillin". <sep>Taboo <sep>Can I take amoxicillin if I have liver problems? The text then inputs the data into a BERT model based on a medical terminology database to predict pharmaceutical entities (such as amoxicillin) and their similar entities in the sentence. Simultaneously, it extracts common-sense keywords related to medicine (such as "eat") and outputs their scores. For example, it outputs the keyword set "{'Amoxicillin': 1.5, 'Contraindications': 1.0, 'Poor liver function': 1.2, 'Eat': 1.2}" and the expanded word set "{'Hydroxybenzylpenicillin': 0.7, 'Amoxicillin granules': 1.1, 'Amoxicillin tablets': 1.0, 'Amoxicillin dry suspension': 0.8, 'Liver or kidney dysfunction': 0.5, 'Poor liver function': 0.9, 'Abnormal liver function': 0.8}". Then, it expands the common-sense keywords using a large language model. For example, it outputs the prompt command "prompt" to generate alternative words containing '[]' in the following sentences and their similarity scores: amoxicillin <sep>[Taboo] <sep>The question "Can I take amoxicillin if I have liver problems?" yields an expanded word set "{'Not suitable': 0.7, 'Contraindications': 0.5, 'Taking': 0.9, 'Ingestion': 0.8, 'Eating': 0.7, 'Digestion': 0.3}", where the score is obtained by multiplying the score of the corresponding keyword by the score of the expanded word generated by the large language model (e.g., similarity). Similarly, the same processing can be performed on any knowledge document in the knowledge base to obtain a keyword set and an expanded word set. For example, for the document knowledge "Contraindications: Patients allergic to amoxicillin and patients with severe liver and kidney dysfunction should not take it to avoid adverse symptoms after medication", the keyword set can be obtained. The input text containing the target sub-problem is combined with the following set of terms: "{'Contraindications': 1.1, 'Population': 0.8, 'Amoxicillin': 1.2, 'Allergy': 0.9, 'Hepatic and Renal Dysfunction': 0.8, 'Usage': 0.9, 'Symptoms': 0.6}". Then, the sparse representations corresponding to the input content containing the target sub-problem text and the knowledge document can be calculated and defined as w respectively. q and w u Then, multiply them by their dot product to obtain the similarity score, which is defined as S. u The system utilizes a large language model to locate knowledge in the K (the specific value is not limited, such as 3) highest-scoring knowledge documents. For example, the knowledge location prompt could be: "Please select the most effective knowledge fragment from the given text to answer the user's question. Output 'None' if no corresponding knowledge is found. User question: [Question] Text: [Text]", where the "[Question]" slot is filled with the search target sub-question text and the searched knowledge, respectively. From the located knowledge with appropriate length, the document with the highest document similarity score is added to the candidate knowledge list.

[0098] In this way, by combining the drug knowledge base and the capabilities of the large language model to construct a sparse representation of the input content containing the target sub-problem text and any knowledge document in the knowledge base, the recall capability of the large language model plugin for unstructured knowledge can be significantly improved through similarity calculation of the sparse representation.

[0099] Furthermore, text filtering rules can be used to filter the retrieved knowledge, deleting information containing only meaningless fields or unclear drug instruction manuals.

[0100] In addition, an alternative implementation is to perform knowledge retrieval using drug instructions and / or a pre-defined drug knowledge base based on the intent classification results of each pharmaceutical entity. If no knowledge is found, a pre-defined web search engine can be used for knowledge searching, and a large language model can be used to locate the knowledge content in the search results. Specifically, the target sub-question text can be input into the pre-defined web search engine (the specific structure is not limited). A knowledge location prompt is used to compare the retrieved knowledge with the searched sub-question using a large language model, outputting the valid portion of the web knowledge. The knowledge location prompt can be the same as or different from the steps described above.

[0101] In this way, after obtaining the knowledge corresponding to each of the N target sub-problems, the knowledge can be arranged and integrated. For example, the knowledge list can be in the form of "{[sub-problem 1]: [knowledge 1], [knowledge 2]; [sub-problem 2]: [knowledge 1], [knowledge 2]}".

[0102] Then, titles can be added to these knowledge items to describe different knowledge types and sources. For the inferences obtained and external knowledge obtained through web search engines, "Sub-problem Conclusion" and "Web Search Knowledge" can be given as knowledge titles, respectively. For knowledge retrieved through the knowledge base, the original pharmaceutical entity and intent tag in the question can be concatenated as the knowledge title. At the same time, the actual pharmaceutical entity name (such as standard term) used for querying after mapping and the intent tag can be concatenated and inserted before the knowledge content to anchor the difference between the actual query object and the user's query object.

[0103] Furthermore, drug type mapping descriptions can be inserted into the prompt, such as "(There are many types of painkillers; acetaminophen tablets, which are commonly used, are used as an example.)". At the same time, the transition from drug type to specific drug can be explained using a large language model.

[0104] In addition, citation numbers can be added to knowledge and the knowledge source of each citation number can be recorded. At the same time, when the large language model uses knowledge from the knowledge list, it outputs the citation number at the corresponding position, so that the basis of the large language model's answer is traceable.

[0105] For example, the knowledge fusion prompt is shown below, where fields marked with "【】" indicate possible text slots:

[0106] "Answer user questions by combining the sub-questions and corresponding knowledge in the knowledge list. When referring to knowledge in the knowledge list, the corresponding knowledge number should be given at the reference location."

[0107] Knowledge List:

[0108]

[0109]

[0110] To illustrate further: when the target drug question text is: "Can I take aspirin if I have liver problems? How much should I take at once?", the knowledge fusion prompt obtained after the above steps is as follows:

[0111] "Answer user questions by combining the sub-questions and corresponding knowledge in the knowledge list. When referring to knowledge in the knowledge list, the corresponding knowledge number should be given at the reference location."

[0112] Knowledge List:

[0113]

[0114]

[0115]

[0116] Then input the prompt command into the preset large language model, and the response content of the large language model to the above target drug question text can be: "For patients with poor liver function, special care should be taken when using aspirin. Aspirin may cause liver function test values ​​to rise during use [1], so it is recommended that patients with liver damage use aspirin with caution [2]. Regarding the usage and dosage of aspirin, taking 100mg / tablet enteric-coated aspirin tablets as an example, the daily dose is 1 tablet when used for unstable angina [3]. However, the specific usage and dosage need to be determined according to your specific situation and the doctor's advice. Before using aspirin, please be sure to consult a doctor and follow the doctor's advice. I hope your liver function can recover to health as soon as possible."

[0117] In summary, the drug knowledge question-answering method provided in this embodiment first obtains the target drug question text to be answered; then decomposes the target drug question text to obtain N target sub-question texts; and extracts pharmaceutical entities from the N target sub-question texts to obtain pharmaceutical entities in the N target sub-question texts; then concatenates the pharmaceutical entities and the N target sub-questions pairwise to obtain each pharmaceutical entity and target sub-question pair, and performs intent classification on each pharmaceutical entity and target sub-question pair to obtain the first intent classification result for each pharmaceutical entity; next, using preset pharmaceutical rules, performs intent classification on the target sub-questions in each pharmaceutical entity and target sub-question pair to obtain the second intent classification result for each pharmaceutical entity; and merges it with the first intent classification result to obtain the intent classification result for each pharmaceutical entity; then, based on the intent classification results of each pharmaceutical entity, knowledge retrieval is performed using the drug instructions and / or a preset drug knowledge base, and the retrieved knowledge, combined with the prompt instruction, is input into a preset large language model to obtain the response content for the target drug question text output by the large language model.

[0118] As can be seen, since this application first decomposes the target drug question text into sub-questions and extracts pharmaceutical entities, it is beneficial to achieve fine-grained classification of the target user's question intent. Then, it integrates the drug knowledge question-answering plugin with the preset pharmaceutical rules and the preset large language model. Based on the intent classification results of each pharmaceutical entity, it uses the drug instructions and / or the preset drug knowledge base to perform more accurate knowledge retrieval, thereby effectively improving the accuracy of subsequent answers to the target drug question and thus improving the interactive experience of the target user.

[0119] Second Embodiment

[0120] This embodiment will introduce a drug knowledge question-and-answer device; please refer to the above method embodiment for related content.

[0121] See Figure 4 This is a schematic diagram of the composition of a drug knowledge question-and-answer device provided in this embodiment. The device 400 includes:

[0122] The acquisition unit 401 is used to acquire the target drug question text to be answered; decompose the target drug question text to obtain N target sub-question texts; and extract pharmaceutical entities from the N target sub-question texts to obtain the pharmaceutical entities in the N target sub-question texts; where N is a positive integer greater than 0.

[0123] The first classification unit 402 is used to concatenate the pharmaceutical entity and the N target sub-problems in pairs to obtain each pharmaceutical entity and target sub-problem pair, and to perform intent classification on each pharmaceutical entity and target sub-problem pair to obtain the first intent classification result of each pharmaceutical entity;

[0124] The second classification unit 403 is used to classify the intentions of the target sub-problems in each pharmaceutical entity and target sub-problem pair using preset pharmaceutical rules, to obtain the second intention classification result of each pharmaceutical entity; and to merge it with the first intention classification result to obtain the intention classification result of each pharmaceutical entity.

[0125] The retrieval unit 404 is used to perform knowledge retrieval based on the intent classification results of each pharmaceutical entity, using the drug instructions and / or a preset drug knowledge base, and input the retrieved knowledge, combined with the prompt instruction, into a preset large language model to obtain the response content of the large language model outputting the question text for the target drug.

[0126] In one implementation of this embodiment, the acquisition unit 401 is specifically used for:

[0127] Combining the sub-problem splitting prompt instruction, the target drug problem text is decomposed using a preset large language model to obtain N target sub-problem texts;

[0128] The large language model is obtained by training language rules and patterns using a large-scale language dataset through autoregressive generation.

[0129] In one implementation of this embodiment, the pharmaceutical entity includes a drug entity, a disease entity, and a symptom entity.

[0130] In one implementation of this embodiment, the second classification unit 403 includes:

[0131] An extraction subunit is used to extract patient information entities from the N target sub-question texts to obtain patient information entities in the N target sub-question texts; and to construct a patient profile using the patient information entities and pharmaceutical entities;

[0132] The comparison subunit is used to analyze and compare patient profiles using reasoning logic pre-built based on doctors' experience, combined with a drug knowledge base, to arrive at a conclusion on whether each drug should be used.

[0133] The determination subunit is used to determine the knowledge recommendation content for the corresponding drug based on the conclusion of whether each drug should be used, and to use the recommended knowledge content as the second intent classification result of the corresponding drug entity.

[0134] In one implementation of this embodiment, the retrieval unit 404 includes:

[0135] Obtain sub-units to use drug instructions and / or a preset drug knowledge base to perform entity alignment and entity disambiguation on each pharmaceutical entity to obtain the corresponding generic drug name, standard disease term, and standard symptom term;

[0136] The query subunit is used to query unstructured, semi-structured, and structured knowledge in drug instructions and / or a preset drug knowledge base based on the intent classification results of each of the pharmaceutical entities, and use this as the retrieval result.

[0137] In one implementation of this embodiment, the query subunit is specifically used for:

[0138] Based on the intent classification results of drug entities, query unstructured knowledge in the drug instructions as the search results;

[0139] And / or, based on the intent classification results of disease entities, query the structured knowledge in the preset drug knowledge base as the search results;

[0140] And / or, based on the intent classification results of each of the pharmaceutical entities, query unstructured knowledge in a pre-defined drug knowledge base using sparse representation as the retrieval result.

[0141] In one implementation of this embodiment, the apparatus further includes:

[0142] The search unit is used to perform knowledge retrieval using drug instructions and / or a preset drug knowledge base based on the intent classification results of each pharmaceutical entity. If no knowledge is found, it performs knowledge search through a preset web search engine and uses the large language model to locate the knowledge content in the search results.

[0143] Furthermore, embodiments of this application also provide a drug knowledge question-and-answer device, including: a processor, a memory, and a system bus;

[0144] The processor and the memory are connected via the system bus;

[0145] The memory is used to store one or more programs, the one or more programs including instructions, which, when executed by the processor, cause the processor to perform any of the above-described implementations of the drug knowledge question-and-answer method.

[0146] Furthermore, embodiments of this application also provide a computer-readable storage medium storing instructions that, when executed on a terminal device, cause the terminal device to perform any of the above-described implementations of the drug knowledge question-and-answer method.

[0147] Furthermore, this application embodiment also provides a computer program product, which, when run on a terminal device, causes the terminal device to execute any of the above-described implementation methods of the drug knowledge question-and-answer method.

[0148] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that all or part of the steps in the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network communication device such as a media gateway, etc.) to execute the methods described in various embodiments or some parts of the embodiments of this application.

[0149] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0150] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0151] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.< / sep> < / sep> < / sep> < / sep> < / sep> < / sep> < / sep>

Claims

1. A drug knowledge question-and-answer method, characterized in that, include: Retrieve the text of the target drug question that is pending a response; The target drug problem text is then decomposed into N target sub-problem texts; And perform pharmaceutical entity extraction on the N target sub-problem texts to obtain pharmaceutical entities in the N target sub-problem texts; where N is a positive integer greater than 0; the pharmaceutical entities include drug entities, disease entities, and symptom entities; The pharmaceutical entity and the N target sub-problems are concatenated in pairs to obtain each pharmaceutical entity and target sub-problem pair. The intent of each pharmaceutical entity and target sub-problem pair is then classified to obtain the first intent classification result of each pharmaceutical entity. Using preset pharmaceutical rules, the intention classification of the target sub-problems in each pair of pharmaceutical entities and target sub-problems is performed to obtain the second intention classification result of each pharmaceutical entity; Then, merge it with the first intent classification result to obtain the intent classification result for each of the pharmaceutical entities; Based on the intent classification results of each pharmaceutical entity, knowledge retrieval is performed using the drug instructions and / or a preset drug knowledge base. The retrieved knowledge is then combined with a prompt instruction and input into a preset large language model to obtain the response content of the large language model for the target drug question text. The step of using preset pharmaceutical rules to classify the intent of each pharmaceutical entity and target sub-problem pair to obtain a second intent classification result for each pharmaceutical entity includes: Patient information entities are extracted from the N target sub-problem texts to obtain patient information entities in the N target sub-problem texts; and patient profiles are constructed using the patient information entities and pharmaceutical entities; By using reasoning logic pre-constructed based on doctors' experience, combined with a drug knowledge base, patient profiles are analyzed and compared to entities to arrive at conclusions on whether each drug should be used. Based on the conclusions regarding whether each drug should be used, the content recommended for the corresponding drug is determined, and the recommended knowledge content is used as the second intent classification result for the corresponding drug entity.

2. The method according to claim 1, characterized in that, The target drug problem text is decomposed into N target sub-problem texts, including: Combining the sub-problem splitting prompt instruction, the target drug problem text is decomposed using a preset large language model to obtain N target sub-problem texts; The large language model is obtained by training language rules and patterns using a large-scale language dataset through autoregressive generation.

3. The method according to claim 2, characterized in that, The step of performing knowledge retrieval based on the intent classification results of each pharmaceutical entity, using drug instructions and / or a pre-defined drug knowledge base, includes: Using drug instructions and / or a pre-set drug knowledge base, entity alignment and disambiguation are performed on each pharmaceutical entity to obtain the corresponding generic drug name, standard disease term, and standard symptom term. Based on the intent classification results of each pharmaceutical entity, the unstructured, semi-structured, and structured knowledge in the drug instructions and / or the preset drug knowledge base are queried as the search results.

4. The method according to claim 3, characterized in that, The step of querying unstructured, semi-structured, and structured knowledge in drug instructions and / or a pre-defined drug knowledge base based on the intent classification results of each pharmaceutical entity, as the retrieval results, includes: Based on the intent classification results of drug entities, query unstructured knowledge in the drug instructions as the search results; And / or, based on the intent classification results of disease entities, query the structured knowledge in the preset drug knowledge base as the search results; And / or, based on the intent classification results of each of the pharmaceutical entities, query unstructured knowledge in a pre-defined drug knowledge base using sparse representation as the retrieval result.

5. The method according to any one of claims 1-4, characterized in that, When, based on the intent classification results of each pharmaceutical entity, a knowledge retrieval is performed using the drug instructions and / or a pre-defined drug knowledge base, and no knowledge is found, the method further includes: Knowledge is searched using a pre-set web search engine, and the knowledge content in the search results is located using the large language model.

6. A drug knowledge question-and-answer device, characterized in that, include: The acquisition unit is used to acquire the text of the target drug question to be answered. The target drug problem text is then decomposed into N target sub-problem texts; And perform pharmaceutical entity extraction on the N target sub-problem texts to obtain pharmaceutical entities in the N target sub-problem texts; where N is a positive integer greater than 0; the pharmaceutical entities include drug entities, disease entities, and symptom entities; The first classification unit is used to concatenate the pharmaceutical entity and the N target sub-problems in pairs to obtain each pharmaceutical entity and target sub-problem pair, and to perform intent classification on each pharmaceutical entity and target sub-problem pair to obtain the first intent classification result of each pharmaceutical entity; The second classification unit is used to classify the intentions of the target sub-problems in each pharmaceutical entity and target sub-problem pair using preset pharmaceutical rules, and to obtain the second intention classification result of each pharmaceutical entity. Then, merge it with the first intent classification result to obtain the intent classification result for each of the pharmaceutical entities; The retrieval unit is used to perform knowledge retrieval based on the intent classification results of each pharmaceutical entity, using the drug instructions and / or a preset drug knowledge base, and inputting the retrieved knowledge, combined with the prompt instruction, into a preset large language model to obtain the response content of the large language model outputting the question text for the target drug. The second classification unit includes: An extraction subunit is used to extract patient information entities from the N target sub-question texts to obtain patient information entities in the N target sub-question texts; and to construct a patient profile using the patient information entities and pharmaceutical entities; The comparison subunit is used to analyze and compare patient profiles using reasoning logic pre-built based on doctors' experience, combined with a drug knowledge base, to arrive at a conclusion on whether each drug should be used. The determination subunit is used to determine the knowledge recommendation content for the corresponding drug based on the conclusion of whether each drug should be used, and to use the recommended knowledge content as the second intent classification result of the corresponding drug entity.

7. A drug knowledge question-and-answer device, characterized in that, include: Processor, memory, system bus; The processor and the memory are connected via the system bus; The memory is used to store one or more programs, the one or more programs including instructions that, when executed by the processor, cause the processor to perform the method according to any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed on a terminal device, cause the terminal device to perform the method described in any one of claims 1-5.

Images