CN110188170B - Multi-entry medical question template device and method thereof - Google Patents

Abstract

The invention discloses a multi-entry medical question template device and a method thereof, wherein the template device is a medical question conversion mechanism based on medical concepts and medical relations, which binds reasoning rules and reasoning functions in a main template structure, a near-meaning template structure and a multi-entry joint structure together, the method and the device realize the conversion of various near-meaning user questions into a multi-entry medical question template, so that corresponding answers are extracted from the UMLS medical knowledge base according to the multi-entry medical question template obtained through conversion. The invention is designed for solving the problems that knowledge and relations are complex and one question meaning needs to be interpreted through various semantic relations in an intelligent information retrieval and automatic question-answering system in the medical industry, and simultaneously, the invention also aims at improving the template design efficiency, reducing the scale of a template library and meeting the question intention of a user to the maximum extent.

Description

A multi-entry medical question template device and method thereof

Technical Field

The present invention belongs to the field of computer artificial intelligence, and in particular, is a multi-entry medical question template device and method thereof, which can be widely used in intelligent information retrieval and automatic question-answering systems in the medical industry.

Background Art

The rapid development of Internet technology has led to an exponential growth in network information. In order to quickly and accurately obtain the required information from this massive amount of information, traditional search engines can no longer meet the needs of users. Intelligent information retrieval and automatic question-answering systems are gradually becoming the focus and hotspot of research because they can accurately and directly answer user questions. However, due to the diversity and randomness of user questions, the accurate acquisition of the semantics of user questions is quite different from the expectations. Therefore, the accuracy of automatic question-answering and intelligent information retrieval answers is not high.

Question processing and understanding is the first execution stage in intelligent information retrieval and automatic question answering systems. Its goal is to enable computers to understand users' questions, obtain users' question intentions, and provide a basis for subsequent information retrieval. Question understanding generally includes lexical analysis, syntactic analysis, and semantic analysis. At present, semantic analysis is the bottleneck of natural language processing. In domain-oriented information retrieval and question answering systems, user queries have many similarities. For example, in an ontology-based knowledge base, many queries are about the attributes of concepts or entities, as well as the relationships between concepts and between attributes. These questions can be extracted into representative question templates based on domain ontology that encapsulate semantic information, thereby effectively avoiding complex lexical analysis, syntactic analysis, and semantic analysis. Therefore, it is of great significance to study question understanding methods based on semantic relationships and question templates.

At present, the question understanding method based on semantic relations and question templates generally adopts a template corresponding to a semantic relation in the domain ontology or domain knowledge base. However, in a medical information system, the knowledge and relations are relatively complex, and the semantics of a question need to be explained through multiple semantic relations. For example, when asking about the relevant symptoms of a disease, multiple semantic relations are needed to express it. A single semantic relation is not comprehensive in expressing the symptoms of the disease and cannot provide a more comprehensive description of the symptoms of the disease. In order to make the semantics of the template more comprehensive, clear and clear, and at the same time to improve the template matching accuracy and design efficiency, and reduce the size of the template library, it is very necessary to find a multi-entry medical question template device and application method based on medical concepts and medical relations.

Summary of the invention

The present invention is designed to solve the problem in intelligent information retrieval and automatic question-answering systems in the medical industry that knowledge and relationships are relatively complex and the semantics of a question need to be explained through multiple semantic relationships. At the same time, it also provides a multi-entry medical question template device and method to improve template design efficiency, reduce the size of the template library, and maximize the satisfaction of users' question intentions.

To achieve the above object, the technical solution of the present invention is:

A multi-entry medical question template device is a medical question conversion mechanism based on medical concepts and medical relationships. It binds together the inference rules and inference functions in the main template structure, the synonymous template structure, and the multi-entry joint structure to realize the conversion of multiple synonymous user questions into a multi-entry medical question template, so as to extract the corresponding answers from the UMLS medical knowledge base according to the multi-entry medical question template obtained by the conversion.

Furthermore, the question template binds the inference rules and inference functions in the main template structure, the synonymous template structure, and the multi-entry joint structure together, and its Backus-Naur form is defined as:

<Multi-entry medical question template>::=(<main template structure>,{<synonymous template structure>},<multi-entry joint structure>) (1)

<Multiple entry union structure>::=({<entry union structure>}) (2)

<Entry Union Structure>::=({<Synonymous Template Structure>}, <Preferred Binding Structure>, {<Secondary Binding Structure>}) (3)

<preferred binding structure>::=(<preferred inference rule>, <preferred inference function>) (4)

<Secondary Binding Structure>::=(<Secondary Inference Rule>, <Secondary Inference Function>) (5)

The main template structure is the most representative sentence structure of the question template represented by variables and labels, reflecting the shallow question semantics of the question template;

The synonymous template structure is a template structure with similar semantics to the main template structure; a question template contains more than one synonymous structure;

The multi-entry joint structure is a mechanism for converting various sub-semantics of a question template into different semantic relationships in a medical knowledge system; a question template contains more than one entry joint structure; each entry joint structure represents a reasoning method for a certain sub-semantic of the question template in which it is located, and is composed of a group of synonymous template structures, preferred reasoning rules and their reasoning functions, and a plurality of ordered secondary reasoning rules and their reasoning functions; wherein the synonymous template structure is a group of synonymous question sentence structures reflecting the sub-semantics of the entry joint structure, each reasoning rule is associated with a corresponding reasoning function, the preferred reasoning rule is the answer reasoning method that best conforms to the template structure semantics of the entry joint structure, and the secondary reasoning rules are sorted according to their closeness to the template structure semantics;

The inference rule is a deep question semantics representing a question template, and uses a predicate formula based on medical concepts and medical relations to accurately represent the reasoning process of the expected answer and the user's intention of asking the question;

The inference function is an answer inference program bound to the inference rule, which executes the inference function specified by the inference rule and extracts the corresponding answer from the medical knowledge base through the semantic relationship specified by the inference rule. Its driving semantics comes from the medical knowledge elements in the user question it matches;

The union of the set of "primary template structure" and the set of "synonymous template structure" in formula (1) is equal to the union of the set of "synonymous template structures" in formula (3).

Furthermore, among the main template structure and the synonymous template structure, the normalized semantic similarity of any two template structures based on core elements, variable types and sequences is less than 1.

Furthermore, the template structure is defined as a set of template elements annotated with template annotation symbols and variable type symbols and arranged in a certain order. The template structure includes a main template structure and a synonym template structure. The question semantics of the question template is represented by the variables and core elements in the template structure, which is defined in Backus-Naur form as:

<Template structure>:=(<Template element 1>, <Template element 2>, ..., <Template element n>) (6)

<Template element>:=(<core element>,<optional element>,<variable>) (7)

<Core element>: = Elements marked with the markers “<”, “>” in the template (8)

<Optional element>: = Elements marked with the markers “[”, “]” in the template (9)

<variable>:=<variable name>+“:”+<variable type> (10)

The template markers include: <> is used to define a required core element in the template; [] is used to define an optional element that can be omitted in the template; {} represents a set of elements in the template; | is used to separate synonyms in the template;

Variable type symbols include: ①Concept: <c:Concept> declares that the template variable c is a medical concept in the UMLS knowledge base; ②Relation: <r:Concept> declares that the template variable r is a medical relationship in the UMLS knowledge base; ③ConceptSet: <s:ConceptSet> declares that the template variable s is a medical concept set in the UMLS knowledge base; ④Type: <t:Type> declares that the template variable t belongs to the Type type in the UMLS knowledge base.

Furthermore, the inference rule is a logical implication defined by a template description logic system, the template description logic system is abbreviated as TDLS, and TDLS is the following tuple:

TDLS::＝(<predicate set>,<operator>)

The predicate is used to declare, identify and determine the medical concepts and medical relations in the question template. In TDLS, there are three types of predicates: unary predicates, binary predicates and ternary predicates. The unary predicate is used to declare the category of medical knowledge to which the template variable belongs, and the binary predicate is used to declare the semantic relationship between two template variables. The ternary predicate is used to declare the domain and range of the medical relationship.

表示全称量词，代表任何个体，运算对象为医学概念；④符号

表示存在量词，代表存在某个个体，运算对象为医学概念；⑤符号“:”：表示医学知识变量的类型定义符，运算对象：左边为模板变量，右边为一阶模板谓词；⑥符号“.”：表示医学概念或实例的关系的引用运算对象为医学概念。The operators include: ① symbol “∧”: represents conjunction operation, represents logical “and”, and the operation object is a predicate or a predicate logic expression; ② symbol “∨”: represents disjunction operation, represents logical “or”, and the operation object is a predicate or a predicate logic expression; ③ symbol

It represents a universal quantifier, representing any individual, and the operation object is a medical concept; ④Symbol

It represents an existential quantifier, representing the existence of an individual, and the operation object is a medical concept; ⑤Symbol “:”: It represents the type definition symbol of a medical knowledge variable, and the operation object: the template variable on the left and the first-order template predicate on the right; ⑥Symbol “.”: It represents the reference operation object of the relationship between medical concepts or instances, which is a medical concept.

The method for answer reasoning using the multi-entry medical question template device described above comprises the following steps:

S1. Establish a multi-entry medical question template library:

S11. Collect the user question sets in the UMLS-based medical question answering system and perform root word restoration on all user questions;

S12. Template the user's question using the template marker and the variable type marker, mark the core elements, optional elements, UMLS concept names and relationship names, and replace each UMLS concept name, concept name set and relationship name with a variable name and variable type marker;

S13. classifying the templated user questions, collecting user questions with similar semantics together, and forming a template structure set of multi-entry medical question templates;

S14. Repeat step S13 until all user questions in the user question set are classified, forming a multi-entry medical question template library including only the template structure set;

S15. According to different sub-semantics of the template structure and in combination with the medical concepts and medical relations in the UMLS knowledge base, the template structure set in each multi-entry medical question template is divided into multiple groups to form a multi-entry joint structure;

S16. Using the template description logic system and combining the medical concepts and medical relations in the UMLS knowledge base, design reasoning rules and reasoning functions that conform to the entry joint structure for each entry joint structure of the multi-entry medical question template library.

S2. Matching the user's question to be matched with the question template in the multi-entry medical question template library:

S21. Preprocess the user questions to be matched as follows: first, restore the user questions to their roots, and use the medical knowledge elements in the UMLS unified medical language system as a dictionary to mark the medical concepts and relationships in the user questions. The marking method is: <medical concept: Concept>, <medical concept set: ConceptSet> and <medical relationship: Relation>. Then, mark the nouns of non-medical concepts and verbs of non-medical relationships in the user questions as core elements, and mark the question words and prepositions as core elements. Finally, the user questions marked with core elements and UMLS medical knowledge elements are obtained.

S22. The preprocessed user question is sequentially subjected to sentence template similarity calculations with the question templates in the multi-entry medical question template library, and the question template with the highest sentence template similarity is confirmed as the question template matching the user question.

S3. The entry joint structure containing the question template with the highest matching degree with the user's question is used as the execution entry of the user's question in the matched multi-entry joint structure, and the reasoning function that conforms to the entry joint structure is executed to complete the reasoning and extraction of the answer.

Furthermore, in step S22, the formula for calculating the sentence template similarity between the user's question and each multi-entry medical question template is:

Among them, User represents the user question, MUTP represents a multi-entry medical question template in the template library, TSS represents the template structure set in the multi-entry medical question template MUTP, including the main template structure and the synonymous template structure in MUTP, TS is any template structure in the template structure set TSS, StruSim(User,TS) represents the structural similarity between the user question User and the template structure TS, and the calculation formula is as follows:

StruSim(User,TS)＝VarSim(User,TS)×KeySim(User,TS) (12)

Among them, VarSim(User, TS) represents the similarity between the variables in the template structure TS and the UMLS medical knowledge elements in the user question User, and the calculation method is as shown in formulas (13) and (14). KeySim(User, TS) represents the similarity between the core elements in the template structure TS and the core elements in the user question User, and the calculation method is as shown in formulas (15) and (16):

表示j的类型与i的类型相同或j的类型被i的类型所包含，m为Key中的任意一个核心元素，Key为模板结构TS中的核心元素集合，n为KE中的任意一个核心元素，KE表示用户问句User中的核心元素集合，VS(i)、KS(m)分别表示模板结构TS中的变量i、核心元素m与用户问句User之间的相似度，sim(m,n)表示基于通用语义词典的归一化的词语语义相似度计算，STH为一个相似度阈值；sim(m,n)的计算公式为：Where i is any variable in Var, Var is a variable set in the template structure TS, j is any UMLS medical knowledge element in UE, UE represents the set of UMLS medical knowledge elements in the user question User, Type(i) and Type(j) represent the types of UMLS medical knowledge to which i and j belong.

Indicates that the type of j is the same as the type of i or the type of j is included in the type of i, m is any core element in Key, Key is the core element set in the template structure TS, n is any core element in KE, KE represents the core element set in the user question User, VS(i) and KS(m) represent the similarity between the variable i and the core element m in the template structure TS and the user question User, sim(m,n) represents the normalized word semantic similarity calculation based on the general semantic dictionary, STH is a similarity threshold; the calculation formula of sim(m,n) is:

Where LCS(m,n) represents the nearest common parent node between core elements m and n in the general semantic dictionary, depth(LCS(m,n)) represents the depth of LCS(m,n) in the general semantic dictionary, and pathLen(m,n) represents the shortest path between core elements m and n in the general semantic dictionary.

The universal semantic dictionary refers to a cross-domain computable dictionary based on a classification structure.

Furthermore, in step S3, executing the inference function conforming to the entry joint structure includes the following sub-steps:

S31. First, execute the preferred reasoning function bound to the preferred reasoning rule. If and only if the execution of the preferred reasoning function fails, start executing the secondary reasoning function and go to step S32. Otherwise, return the answer extraction result of the preferred reasoning function and end the execution of the reasoning function.

S32. Execute the first secondary reasoning function. If and only if the first secondary reasoning function fails to execute, start executing the second secondary reasoning function. In this way, execute all the secondary reasoning functions bound to the template in turn. If all reasoning functions fail to execute, return a failure message. Otherwise, return the answer extraction result of the successfully executed reasoning function and end the execution of the reasoning function. The failure of the reasoning function execution means that the reasoning function does not find a record with the specified semantic relationship of the question concept in the UMLS knowledge base.

The multi-entry medical question template device and application method based on medical concepts and medical relationships described above, the multi-entry joint template obtained by the device is a question formula based on medical concepts and medical relationships, which binds the reasoning rules and reasoning functions in the main template structure, the synonymous template structure, and the multi-entry joint structure together to form a conversion from multiple synonymous user questions to multiple similar answers, thereby improving the template matching accuracy and template design efficiency, reducing the size of the template library, and maximizing the satisfaction of the user's question intentions.

The medical knowledge base involved in the multi-entry medical question template of the present invention complies with the UMLS (Unified Medical Language System) unified medical language system proposed by the U.S. National Library of Medicine, that is, the multi-entry medical question template of the present invention is based on the medical concepts and medical relationships of UMLS. UMLS is a medical concept semantic network, in which there are more than one million medical concepts, 133 semantic types of concept types, and more than 76 relationships are established between these concept types, which are then organized in a tree structure. Since the question template of the present invention complies with the requirements of UMLS, it can be used to extract corresponding answers from the UMLS medical knowledge base, thereby maximally satisfying the user's question intention.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a schematic diagram of the application method of the present invention.

DETAILED DESCRIPTION

The present invention is further described below in conjunction with specific embodiments, but the protection scope of the present invention is not limited to the following embodiments.

(I) Definition and structure of multi-entry medical question template

A multi-entry medical question template device, the template device is a multi-entry medical question conversion mechanism based on medical concepts and medical relationships, which binds the inference rules and inference functions in the main template structure, the synonymous template structure, and the multi-entry joint structure together to realize the conversion of multiple synonymous user questions into a multi-entry medical question template, so as to extract the corresponding answer from the UMLS medical knowledge base according to the converted multi-entry medical question template. UMLS refers to the Unified Medical Language System.

Furthermore, the question template binds the inference rules and inference functions in the main template structure, the synonymous template structure, and the multi-entry joint structure together, and its Backus-Naur form (BNF) is defined as:

<Multi-entry medical question template>::=(<main template structure>,{<synonymous template structure>},<multi-entry joint structure>) (1)

<Multiple entry union structure>::=({<entry union structure>}) (2)

<Entry Union Structure>::=({<Synonymous Template Structure>}, <Preferred Binding Structure>, {<Secondary Binding Structure>}) (3)

<preferred binding structure>::=(<preferred inference rule>, <preferred inference function>) (4)

<Secondary Binding Structure>::=(<Secondary Inference Rule>, <Secondary Inference Function>) (5)

Main template structure: The most representative sentence structure of the question template represented by variables and labels, reflecting the shallow question semantics of the question template.

Synonymous template structure: a template structure with similar semantics to the main template structure; a question template can contain multiple synonymous structures. The main template structure and the synonymous template structure are used for template matching.

Multi-entry joint structure: a mechanism for converting various sub-semantics of question templates into different semantic relationships in medical knowledge systems; a question template contains more than one entry joint structure. Each entry joint structure represents the reasoning method of a certain sub-semantic of the question template in which it is located, and is composed of a set of synonymous template structures, preferred reasoning rules and their reasoning functions, and multiple ordered secondary reasoning rules and their reasoning functions. Among them, the synonymous template structure is a set of synonymous question sentence structures that reflect the sub-semantics of the entry joint structure. Each reasoning rule is associated with a corresponding reasoning function, among which the preferred reasoning rule is the answer reasoning method that best conforms to the template structure semantics of the entry joint structure, and the secondary reasoning rules are sorted according to their degree of closeness to the template structure semantics;

Inference rules: represent the deep semantics of the question template, using predicate formulas based on medical concepts and medical relations to accurately express the reasoning process of the expected answer and the user's intention of asking the question.

Reasoning function: an answer reasoning program bound to the reasoning rules, which executes the reasoning function specified by the reasoning rules and extracts the corresponding answer from the medical knowledge base through the semantic relationship specified by the reasoning rules. Its driving semantics comes from the medical knowledge elements in the user questions it matches, and the driving semantics completes the mapping of the formal parameters of the reasoning function to the actual parameters.

(II) Relationships and constraints between template structures

The union of the "primary template structure" and the set of "synonymous template structures" in formula (1) is equal to the union of all the "synonymous template structures" in formula (3), that is, the "primary template structure" or "synonymous template structure" given in formula (1) appears in at least one of the "synonymous template structures" in formula 3.

In the main template structure and the synonymous template structure, the core elements, variable types and their order in any two template structures cannot be exactly the same, and the normalized semantic similarity based on the core elements, variable types and order of any two template structures must be less than 1.

(III) Question template structure based on variables and core elements

(1) Marking symbols

The present invention designs a set of template element markers for the question template structure, which are used to separate and define various elements in the template, as shown in Table 1.

Table 1 Template annotation symbols

Label explain <> Defines a required core element in the template [] Defines an optional element that can be omitted in a template {} Represents a collection of elements in a template | Used to separate synonyms in templates

(2) Variable type symbol

It is used to define the variable type in the template structure, as shown in Table 2. For example, <c ₁ :Concept> indicates that the variable c ₁ is an ontology concept.

Table 2 Variable type symbols in the question template structure

Variable type specifier describe Concept <c:Concept> declares that the template variable c is a medical concept in the UMLS knowledge base Relation <r:Concept> declares that the template variable r is a medical relationship in the UMLS knowledge base ConceptSet <s:ConceptSet> declares that the template variable s is a concept set in the UMLS knowledge base Type <t:Type> declares that the template variable t belongs to the Type type in the UMLS knowledge base

(3) Template structure

The template structure is defined as a set of template elements annotated with the template annotation symbols shown in Table 1 and the variable type symbols shown in Table 2 and arranged in a certain order. The template structure includes a main template structure and a synonym template structure. The question semantics of the question template is represented by the variables and core elements in the template structure, and is defined in Backus-Naur Form (BNF) as follows:

<Template structure>:=(<Template element 1>, <Template element 2>, ..., <Template element n>) (6)

<Template element>:=(<core element>, <optional element>, <variable>) (7)

<Core element>: = Elements marked with the markers “<”, “>” in the template (8)

<Optional element>: = Elements marked with the markers “[”, “]” in the template (9)

<variable>:=<variable name>+“:”+<variable type> (10)

(IV) Inference rules based on predicate logic

In order to accurately express the semantics of multi-entry medical question templates and define the inference rules in the multi-entry question templates, the present invention designs a template logic system. The logic system is a special first-order description logic that uses medical concepts and medical relationships as operation objects, performs semantic operations and semantic interpretations on multi-entry medical question templates, and defines the inference rules as a logical implication defined by the template description logic system. The present invention defines the template description logic system (TDLS) applied to the multi-entry medical question template as the following two-tuple:

TDLS::＝(<predicate set>,<operator>)

(1) Predicates are used to declare, identify and determine medical concepts and medical relationships in question templates; in TDLS, there are three types of predicates: unary predicates, binary predicates and ternary predicates. Unary predicates are used to declare the category of medical knowledge to which the template variable belongs. Binary predicates are used to declare the semantic relationship between two template variables. Ternary predicates are used to declare the domain and range of medical relationships. Among them, unary predicates can also be used as variable type symbols in template structure annotations. Tables 3, 4 and 5 respectively list and explain the unary predicates, binary predicates and ternary predicates in the template description logic of the present invention.

Table 3 Unary predicates in template logic

Unary predicate describe Concept(c) Declare that the template variable c is a concept in medical knowledge Relation(r) Declare that the template variable r is a relation in medical knowledge Answer(s) Declare that the template variable s is a concept set in the medical knowledge base that conforms to the inference rules ConceptSet(s) Declare that the template variable s is a concept set in medical knowledge Type(t) Declare that the template variable t belongs to the Type type in the UMLS knowledge base

Table 4 Binary predicates in template logic

Binary Predicate describe SubClassOf(c ₁ ,c ₂ ) Indicates that concept _c1 is a direct subclass of _c2 PosterityOf(c ₁ ,c ₂ ) Indicates that concept _c1 is a descendant of _c2 EquivalenceClass(c ₁ ,c ₂ ) Indicates that concepts _c1 and _c2 are equivalent classes

Table 5 Ternary predicates in template logic

(2) Operators: This invention further improves the expressiveness of template logic by extending conventional operators of first-order predicates, which are divided into three types of operators: unary, binary, and ternary. Table 6 lists the operators extended by this invention for template logic.

Table 6 Template operators

(V) The medical knowledge base (including medical concepts and medical relationships) used in this invention complies with the standards

The medical knowledge base involved in the multi-entry medical question template of the present invention is in accordance with the UMLS (Unified Medical Language System) proposed by the National Library of Medicine of the United States, that is, the multi-entry medical question template of the present invention is based on the medical concepts and medical relationships of UMLS. UMLS is a medical concept semantic network, in which there are more than one million medical concepts, 133 semantic types of concepts, and more than 76 relationships are established between these concept types, and then organized in a tree structure.

(VI) Application method of multi-entry medical question template device

The application methods include:

S1. First, establish a multi-entry medical question template library. The multi-entry medical question template library of the present invention refers to a collection of multi-entry medical question templates established in a medical question answering system based on UMLS, which reflects the user's interest and demand for medical knowledge based on UMLS. The specific process of establishing a multi-entry medical question template library is as follows:

S11. Collect the user question sets in the UMLS-based medical question answering system and perform root word restoration on all user questions;

S12. Template the user's question using template annotations and variable type annotations, annotate the core elements, optional elements, UMLS concept names and relationship names, and replace each UMLS concept name, concept name set and relationship name with a variable name and variable type annotation;

S13. classifying the templated user questions, collecting user questions with similar semantics together, and forming a template structure set of multi-entry medical question templates;

S14. Repeat step S13 until all user questions in the user question set are classified, forming a multi-entry medical question template library including only the template structure set;

S15. According to different sub-semantics of the template structure and in combination with the medical concepts and medical relations in the UMLS knowledge base, the template structure set in each multi-entry medical question template is divided into multiple groups to form a multi-entry joint structure;

S16. Using the template description logic system (TDLS), combined with the medical concepts and medical relations in the UMLS knowledge base, design reasoning rules and reasoning functions that conform to the entry joint structure for each entry joint structure of the multi-entry medical question template library.

S2. Match the user question to be matched with the question template in the multi-entry medical question template library:

S21. The user questions to be matched are preprocessed as follows: first, the user questions are rooted, and the medical knowledge elements in the UMLS unified medical language system are used as the dictionary to mark the medical concepts and relations in the user questions. The marking method is: <medical concept: Concept>, <medical concept set: ConceptSet> and <medical relationship: Relation>, such as: <chronic podopompholyx: Concept>, then the nouns of non-medical concepts and verbs of non-medical relations in the user questions are marked as core elements, and the question words and prepositions are marked as core elements, and finally the user questions marked with core elements and UMLS medical knowledge elements are obtained; UMLS medical knowledge elements refer to medical concepts, medical concept sets and relations in the UMLS unified medical language system;

S22. The preprocessed user question is sequentially subjected to sentence template similarity calculations with the question templates in the multi-entry medical question template library, and the question template with the highest sentence template similarity is confirmed as the question template matching the user question.

S3. The entry joint structure containing the question template with the highest matching degree with the user's question is used as the execution entry of the user's question in the matched multi-entry joint structure, and the reasoning function that conforms to the entry joint structure is executed to complete the reasoning and extraction of the answer.

Furthermore, in step S22, the formula for calculating the sentence template similarity between the user's question and each multi-entry medical question template is:

Among them, User represents the user question, MUTP represents a multi-entry medical question template in the template library, TSS represents the template structure set in the multi-entry medical question template MUTP, including the main template structure and the synonymous template structure in MUTP, TS is any template structure in the template structure set TSS, StruSim(User,TS) represents the structural similarity between the user question User and the template structure TS. The question template with the maximum value of match(User,MUTP) is the one with the highest matching degree with the user question.

The calculation formula of StruSim(User,TS) is as follows:

StruSim(User,TS)＝VarSim(User,TS)×KeySim(User,TS) (12)

Among them, VarSim(User, TS) represents the similarity between the variables in the template structure TS and the UMLS medical knowledge elements in the user question User, and the calculation method is as shown in formulas (13) and (14). KeySim(User, TS) represents the similarity between the core elements in the template structure TS and the core elements in the user question User, and the calculation method is as shown in formulas (15) and (16):

Among them, i is any variable in Var, Var is the variable set in the template structure TS, j is any UMLS medical knowledge element in UE, UE represents the UMLS medical knowledge element set in the user question User, Type(i)Type(j) represents the UMLS medical knowledge element to which i and j belong.

表示j的类型与i的类型相同或j的类型被i的类型所包含，m为Key中的任意一个核心元素，Key为模板结构TS中的核心元素集合，n为KE中的任意一个核心元素，KE表示用户问句User中的核心元素集合，VS(i)、KS(m)分别表示模板结构TS中的变量i、核心元素m与用户问句User之间的相似度，表示sim(m,n)表示基于通用语义词典的归一化的词语语义相似度计算，通用语义词典是指基于分类结构、跨领域的可计算词典，如美国普林斯顿大学的WordNet或中科院的HowNet；STH为一相似度阈值，本实施例该阈值取0.85。sim(m,n)的计算公式为：, type,

Indicates that the type of j is the same as the type of i or the type of j is included in the type of i, m is any core element in Key, Key is the core element set in the template structure TS, n is any core element in KE, KE represents the core element set in the user question User, VS(i) and KS(m) respectively represent the similarity between the variable i and the core element m in the template structure TS and the user question User, sim(m,n) represents the normalized semantic similarity calculation of words based on the universal semantic dictionary, the universal semantic dictionary refers to a computable dictionary based on a classification structure and across fields, such as WordNet of Princeton University in the United States or HowNet of the Chinese Academy of Sciences; STH is a similarity threshold, which is 0.85 in this embodiment. The calculation formula of sim(m,n) is:

Among them, LCS(m,n) represents the nearest common parent node between core elements m and n in the universal semantic dictionary, depth(LCS(m,n)) represents the depth of LCS(m,n) in the universal semantic dictionary, and pathLen(m,n) represents the shortest path between core elements m and n in the universal semantic dictionary.

Furthermore, in step S3, executing the inference function conforming to the entry joint structure includes the following sub-steps:

S31. First, execute the preferred reasoning function bound to the preferred reasoning rule. If and only if the execution of the preferred reasoning function fails, start executing the secondary reasoning function and go to step S32. Otherwise, return the answer extraction result of the preferred reasoning function and end the execution of the reasoning function.

S32. Execute the first secondary reasoning function. If and only if the first secondary reasoning function fails to execute, start executing the second secondary reasoning function. In this way, execute all secondary reasoning functions bound to the template in sequence. If all reasoning functions fail to execute, return a failure message. Otherwise, return the answer extraction result of the successfully executed reasoning function and end the execution of the reasoning function.

The failure of executing the reasoning function means that the reasoning function does not find a record of the question concept having a prescribed semantic relationship in the UMLS knowledge base.

(VII) Specific examples of multi-entry medical question templates:

In this example, the structure of the multi-entry medical question template proposed by the present invention and its definition method are demonstrated through the actual definition of two multi-entry medical question templates.

Example 1.

Question template 1: Asking about the parts of a concept

Involved relationship: part of relationship (id = 123005000)

Is-a relationship (id=116680003)

<main template structure>::=<what><do>[the]<c:body structure><include|contain>[？ ]

<synonymous template structure>::={<what><category|type><of><c:Concept><be><include|contain>[? ],

<what><be>[the]<category><of><c:Concept>[? ],

<what><be>[the]<part><of><c:body structure>[? ]}

<Entry Union Structure 1>::=(

<Synonymous template structure>::={<what><do>[the]<c:body structure><include|contain>[? ],

<what><be>[the]<part><of><c:body structure>[? ]}

<Preferred reasoning function>::=Reasoning_function1(c,pa)

<Secondary Reasoning Function>::=Reasoning_function1(c,ia)

)

<Entry Union Structure 2>::=(

<Synonymous template structure>::={<what><category|type><of><c:Concept><be><include|contain>[? ],

<what><be>[the]<category|type><of><c:Concept>[? ]}

<Preferred function call>::=Reasoning_function1(c,ia)

<Secondary function call>::=Reasoning_function1(c,pa)

)

<Example>: What does the Entire facial bone include?

What are the types of Colds?

Example 2.

Question template 2: Asking about disease symptoms

Related relations: Associated morphology (id=116676008)

Associated with(id=47429007)

Due to reason (id=42752001)

<Main template structure>::＝<what><be>[the]<obvious><indication><for|of><c:disease>[？ ]

<synonymous template structure>::＝{<what><do><c:disease><relate|connect|associate><to>[？],

<what><be>[the]<obvious><indication><of|for><c:disease>[? ],

<what><be>[the]<obvious><indication><for|of><patients><with><c:disease>[? ],

<what><be>[the]<obvious><symptom><for|of><c:disease>[? ],

<what><cause>[the]<disease><of><c:disease>[? ],

<What><do><c:Concept><relate|connect|associate>to[? ]}

<Entry Union Structure 1>::=(

<Synonymous template structure>::＝{<what><be>[the]<obvious><indication><for|of><c:Concept>[？]

<what><be>[the]<obvious><indication><for|of><patient><with><c:Concept>[? ],

<what><be>[the]<obvious><indication><of|for><c:Concept>[? ],

<what><be>[the]<obvious><symptom><for|of><c:Concept>[? ]}

<Preferred reasoning function>::=Reasoning_function2(c,am)

<Secondary Reasoning Function 1>::=Reasoning_function2(c,aw)

<Secondary Reasoning Function 2>::＝Reasoning_function2(c,dt)

)

<Entry Union Structure 2>::=(

<synonymous template structure>::＝<what><cause>[the]<disease><of><c:disease>[？ ]

<Preferred reasoning function>::=Reasoning_function2(c,dt)

<Secondary Reasoning Function 1>::=Reasoning_function2(c,aw)

<Secondary Reasoning Function 2>::=Reasoning_function2(c,am)

)

<Entry Union Structure 3>::=(

<Synonymous template structure>::＝<what><do><c:disease><relate|connect|associate><to>[？]

<Preferred reasoning function>::=Reasoning_function2(c,aw)

<Secondary Reasoning Function 1>::=Reasoning_function2(c,am)

<Secondary Reasoning Function 2>::＝Reasoning_function2(c,dt)

)

<Example sentence>: What are the symptoms for Staphylococcus epidermidismeningitis?

What causes the disease of Diabetic maculopathy?

(VIII) Application examples of multi-entry medical question templates

In this example, two medical question-answering systems based on SNOMED-CT are used to demonstrate the application effect of the multi-entry medical question template proposed in the present invention. SNOMED-CT (Systematized Nomenclature of MedicineClinical Terms) is a concept-based structured comprehensive clinical terminology set developed and maintained by the International Health Terminology Standards Development Organization (IHTSDO). The concepts of UMLS come from more than one hundred different source vocabularies, and SNOMED-CT is the most important source vocabulary of UMLS, that is, the SNOMED-CT ontology is the most important subset of UMLS.

SNOMED-CT medical ontology is a concept hierarchy system, in which there is a very special root concept, which contains all the concept hierarchies in SNOMED-CT. In SNOMED-CT, the concepts directly connected to the root concept through the "is-a" relationship are called top-level concepts, and the remaining concepts are connected to at least one top-level concept through the "is-a" relationship. These top-level concepts have different meanings in SNOMED-CT. All top-level concept IDs and meanings are shown in Table 7 below:

Table 7 Top-level concepts in SNOMED-CT

The medical semantic relations that can be formed by concepts in SNOMED-CT are shown in Table 8.

Table 8 Semantic relationships of traditional Chinese medicine in SNOMED-CT

SNOMED-CT is a huge medical knowledge base. Today, it contains more than 440,000 active medical concepts and more than 1 million active medical relationship records. This example uses SNOMED-CT as the knowledge base, and implements two medical question answering systems using the traditional single-entry, single-execution route medical question template and the multi-entry, multi-execution route medical question template proposed in the present invention. The superiority of the multi-entry, multi-execution route medical question template proposed in the present invention in the medical question answering system is verified by comparison. The experimental results are shown in Table 9:

Table 9 Performance comparison of two medical question answering systems based on SNOMED-CT ontology

Question answering system Question Template Type Question template library size Customer satisfaction System 1 Single entry, single execution route 150 78% System 2 Multiple entry points and multiple execution routes 60 85%

The calculation formula of user satisfaction is as follows:

Table 9 shows that the question answering system 2 based on the template of the present invention has greatly reduced the number of question templates of the question answering system 1 based on the traditional template from 150 to 60, thereby improving the efficiency of template design. This is mainly due to the fact that the multi-entry medical template of the present invention has concentrated the question structures with synonymous or near-synonymous question semantics into one template and processed them separately through multiple execution routes. On the other hand, Table 9 also shows that the question answering system 2 based on the template of the present invention has greatly improved the user satisfaction of the question answering system 1 based on the traditional template from 78% to 85%, which shows that the multi-entry medical template of the present invention can significantly improve the performance of the medical question answering system. This is mainly due to the fact that the multi-entry medical template of the present invention uses multiple bound reasoning functions to process a matching question structure, and each reasoning function attempts to infer the answer through a certain semantic relationship. Therefore, when the question concept does not have the semantic relationship specified by the preferred reasoning function, the system will try to obtain the answer through the remaining reasoning functions, thereby expanding the reasoning range of the answer.

Claims (4)

1. A multi-entry medical question template device is characterized in that the template device is a medical question conversion mechanism based on medical concepts and medical relations, and binds reasoning rules and reasoning functions in a main template structure, a near-meaning template structure and a multi-entry joint structure together to realize the conversion of multiple near-meaning user questions into a multi-entry medical question template so as to extract corresponding answers from a UMLS medical knowledge base according to the multi-entry medical question template obtained by conversion;

the question template binds the reasoning rules and the reasoning functions in the main template structure, the near-meaning template structure and the multi-entry joint structure, and the Bakes-Norwalk is defined as follows:

< multiple entry medical question template >: = (< master template structure >, { < near-sense template structure > }, < multiple entry association structure >) (1)

< multiple inlet joint Structure >: = ({ < inlet joint Structure > }) (2)

< entry association structure >: = ({ < synonymous template structure > }, < preferred binding structure >, { < secondary binding structure > }) (3)

< preferred binding Structure >: = (< preferred inference rule >, < preferred inference function >) (4)

< secondary binding structure >: = (< secondary inference rule >, < secondary inference function >) (5)

The main template structure is the most representative sentence structure of a question template expressed by using variables and labels, and reflects shallow question semantics of the question template;

the near-sense template structure is a template structure with similar semantics to the main template structure; one question template comprises more than one near-meaning structure;

the multi-entry joint structure is a conversion mechanism for realizing the conversion from various sub-semantics of the question template to different semantic relations in the medical knowledge system; one question template comprises more than one entry association structure; each entry joint structure represents a certain sub-semantic reasoning method of a question template where the entry joint structure is located, and the method consists of a group of synonymous template structures, preferred reasoning rules and reasoning functions thereof, and a plurality of ordered sub-choice reasoning rules and reasoning functions thereof; the system comprises a plurality of input combined structure sub-semantics, a plurality of first-choice inference rules, a plurality of second-choice inference rules, a plurality of first-choice inference rules and a plurality of second-choice inference rules, wherein the first-choice inference rules are a group of synonymous question sentence structures reflecting the input combined structure sub-semantics, each inference rule is associated with a corresponding inference function, the first-choice inference rules are answer inference methods which are most consistent with the input combined structure template structure semantics, and the second-choice inference rules are ordered according to the similarity degree of the second-choice inference rules and the template structure semantics;

The reasoning rule is a deep question meaning representing a question template, and a predicate formula based on medical concepts and medical relations is used for accurately representing the reasoning process of the expected answer and the intention of the user to question;

the reasoning function is an answer reasoning program bound with the reasoning rules, executes the reasoning function appointed by the reasoning rules, extracts corresponding answers from the medical knowledge base through semantic relations appointed by the reasoning rules, and drives the semantics to come from medical knowledge elements in the matched user questions;

the union of the main template structure and the near-sense template structure in the formula (1) is equal to the union of the synonymous template structures in the formula (3);

in the main template structure and the near-meaning template structure, the normalized semantic similarity of any two template structures based on core elements, variable types and sequences is less than 1;

the template structure is defined as a group of template elements which are marked by using a template mark symbol and a variable type symbol and are arranged in a certain sequence, the template structure comprises a main template structure and a near-meaning template structure, the question semantics of the question template are characterized by the variables and core elements in the template structure, and the template structure is defined as follows by the Barceis-North paradigm:

< template Structure > = (< template element 1>, < template element 2>, …, < template element n >) (6)

< template element > = (< core element >, < optional element >, < variable >) (7)

< core element > = element (8) labeled with the labels "<", ">" in the template

Optional element: = element (9) labeled with the label "[", "]" in the template

< variable > = < variable name > + ": "+ < variable type > (10)

Wherein the template label comprises: defining an optional core element in the template; [] An optional element for defining an eligibility in the template, { } represents a set of elements in the template; the i is used to separate synonyms in templates;

the variable type identifier includes: (1) concept: < c > Concept > states that template variable c is a medical Concept in the UMLS knowledge base; (2) and (3) Relation: < r > Concept > states that the template variable r is a medical relationship in the UMLS knowledge base; (3) ConceptSet: < s ConceptSet > states that the template variable s is a set of medical concepts in the UMLS knowledge base; (4) type: < t: type > states that the template variable t belongs to the Type in the UMLS knowledge base;

The inference rule is a logic implication formula defined by a template description logic system, wherein the template description logic system is called TDLS for short, and the TDLS is the following binary group:

TDLS = (< predicate set >, < operator >)

The predicate is used for declaring, identifying and determining medical concepts and medical relations in the question template; in TDLS, three types of predicates are included: a unitary predicate, a binary predicate, and a ternary predicate; the binary predicate is used for declaring the category of medical knowledge to which the template variable belongs, and the binary predicate is used for declaring the semantic relation between two template variables; the ternary predicates are used for declaring the definition domain and the value domain of the medical relation;

the operators include: (1) symbol ". Lambda": representing conjunctive operation, representing logic AND, wherein an operation object is predicate or predicate logic expression; (2) symbol "[ V-shaped": representation disjunctive operation, representing logic OR, wherein an operation object is predicate or predicate logic expression; (3) sign symbol

Representing the full-scale word, representing any individual, and the operation object is a medical concept; (4) sign->

Representing the presence of a graduated word, representing the presence of a certain individual, and the operation object is a medical concept; (5) symbols ": representing medical knowledge Type definitions of identification variables, operand: the left side is a template variable, and the right side is a first-order template predicate; (6) sign "": a referencing operand representing a relationship of a medical concept or instance is a medical concept.

2. A method for answer reasoning using the multi-entry medical question template apparatus of claim 1, comprising the steps of:

s1, establishing a multi-entry medical question template library:

s11, collecting a user question set in a UMLS-based medical question-answering system, and performing root reduction on all user questions;

s12, templating a user question by using the template identifier and the variable type identifier, marking the core element, the optional element and the concept names and the relationship names of UMLS in the template identifier and the variable type identifier, and replacing each UMLS concept name, each UMLS concept name set and each UMLS relationship name by using a variable name and a variable type identifier respectively;

s13, classifying the user questions after the templated, and collecting the user questions with similar semantics together to form a template structure set of the multi-entry medical question template;

s14, repeating the step S13 until all user question sentences in the user question sentence set are classified, and forming a multi-entry medical question sentence template library only comprising a template structure set;

S15, according to different sub-semantics of the template structure, combining medical concepts and medical relations in the UMLS knowledge base, dividing a template structure set in each multi-entry medical question template into a plurality of groups to form a plurality of entry joint structures;

s16, describing a logic system by using a template, and simultaneously designing an inference rule and an inference function which accord with the entrance joint structure for each entrance joint structure of the multi-entrance medical question template library by combining medical concepts and medical relations in the UMLS knowledge base;

s2, matching the user question to be matched with a question template in the multi-entry medical question template library:

s21, preprocessing a user question to be matched as follows: firstly, performing root reduction on a user question, and marking medical concepts and relations in the user question by taking medical knowledge elements in a UMLS unified medical language system as a dictionary in the following marking modes: < medical concept: receive >, < set of medical concepts: conceptSet > vs < medical relationship: the Relation is carried out, nouns of non-medical concepts and verbs of non-medical relations in user question sentences are marked as core elements, query words and prepositions are marked as core elements, and finally user question sentences marked with the core elements and UMLS medical knowledge elements are obtained;

S22, carrying out sentence pattern similarity calculation on the preprocessed user question and question templates in a multi-entry medical question template library in sequence, and confirming the question template with the highest sentence pattern similarity as a question template matched with the user question;

s3, taking an entrance joint structure where a question template with highest matching degree with a user question is located as an execution entrance of the user question in the matched multi-entrance joint structure, and executing an inference function conforming to the entrance joint structure to complete the inference and extraction of answers.

3. The method according to claim 2, characterized in that:

in the step S22, the sentence pattern similarity calculation formula of the user question and each multi-entry medical question template is:

wherein, user represents a User question, MUTP represents a multi-entry medical question template in a template library, TSS represents a template structure set in the multi-entry medical question template MUTP, the multi-entry medical question template comprises a main template structure and a near-meaning template structure in the MUTP, TS is any one template structure in the template structure set TSS, struSim (User, TS) represents the structural similarity between the User question User and the template structure TS, and the calculation formula is as follows:

StruSim(User,TS)＝VarSim(User,TS)×KeySim(User,TS) (12)

wherein VarSim (User, TS) represents the similarity between the variable in the template structure TS and the UMLS medical knowledge element in the User question User, the calculation methods are shown in formulas (13) and (14), keySim (User, TS) represents the similarity between the core element in the template structure TS and the core element in the User question User, and the calculation methods are shown in formulas (15) and (16):

Wherein i is any one variable in Var, var is a variable set in a template structure TS, j is any one UMLS medical knowledge element in UE, UE represents a UMLS medical knowledge element set in a User question User, type (i) and Type (j) represent types of UMLS medical knowledge to which i and j belong,

the type of j is the same as the type of i or the type of j is contained by the type of i, m is any core element in Key, key is a core element set in a template structure TS, n is any core element in KE, KE is a core element set in a User question User, VS (i) and KS (m) respectively represent the similarity between a variable i, the core element m and the User question User in the template structure TS, sim (m, n) represents normalized word semantic similarity based on a general semantic dictionaryCalculating the degree, wherein STH is a similarity threshold; the calculation formula of sim (m, n) is: />

Wherein LCS (m, n) represents the nearest common parent node in the generic semantic dictionary between core elements m, n, depth (LCS (m, n)) represents the depth of LCS (m, n) in the generic semantic dictionary, pathLen (m, n) represents the shortest path in the generic semantic dictionary between core elements m, n;

the universal semantic dictionary refers to a cross-domain computable dictionary based on a classification structure.

4. The method according to claim 2, characterized in that:

in the step S3, executing the inference function conforming to the portal federation structure includes the following sub-steps:

s31, firstly executing the first-choice reasoning function bound with the first-choice reasoning rule, if and only if the execution of the first-choice reasoning function fails, starting to execute the second-choice reasoning function, and turning to the step S32, otherwise, returning an answer extraction result of the first-choice reasoning function and ending the execution of the reasoning function;

s32, executing a first sub-selection reasoning function, if and only if the first sub-selection reasoning function fails to be executed, starting to execute a second sub-selection reasoning function, sequentially executing all sub-selection reasoning functions bound by the template in the mode, returning failure information if all the reasoning functions fail to be executed, otherwise, returning an answer extraction result of the successfully executed reasoning function and ending the execution of the reasoning function; the reasoning function execution failure means that the reasoning function does not find that the questioning concept has a specified semantic relation record in the UMLS knowledge base.

Images