CN110609907A - A random walk-based reasoning method for medical domain knowledge - Google Patents

Abstract

The invention relates to a reasoning method for medical field knowledge based on random walk. The content of the invention mainly includes (1) a named entity recognition method in the medical field based on contextual character binary groups and information entropy; (2) a method for extracting relationships between entities in the medical field based on predicate sentiment classification; (3) a medical field based on random walk Domain knowledge graph reasoning method. Based on the above method, the named entities in the medical field are identified, and the relationship between named entities is extracted, so as to automatically construct the knowledge map in the medical field and realize the reasoning of the knowledge map in the medical field.

Description

A random walk-based reasoning method for medical domain knowledge

technical field

The invention relates to the fields of knowledge engineering and machine learning, and relates to a random walk-based reasoning method for medical field knowledge.

Background technique

As one of the key technologies in the field of knowledge engineering and artificial intelligence, knowledge graph technology is one of the current hot technical research fields. Different from machine learning technology, there are often problems that it is difficult to explain the local relationship between features and the global relationship between features and output. Knowledge graph technology uses triples to represent the relationship between knowledge entities, which intuitively reflects the knowledge ontology and the association logic between knowledge entities. , has very good explainability, has been paid more and more attention by the industry, and has become one of the important foundations of artificial intelligence technology.

Knowledge map technology mainly includes construction, reasoning, etc. Among them, knowledge map construction technology mainly includes named entity recognition, relationship extraction, etc., knowledge map reasoning technology mainly includes entity relationship prediction, knowledge reasoning, etc. By identifying knowledge entities from the facts in the text data, extracting the relationship between knowledge entities, constructing a knowledge graph based on the triple representation, and completing the knowledge graph by mining and predicting possible relationships between entities, based on The relationship between known entities in the knowledge graph is used to extract and reason knowledge rules.

As a knowledge-intensive field, the medical field relies heavily on the background knowledge of medicine and pharmacy. The use of knowledge graphs to represent the background knowledge of medicine and pharmacy plays a very important role in supporting the application of auxiliary intelligence in the field of medicine. However, named entities, relationships between entities, and knowledge logic in the medical field have very distinct domain characteristics, which are quite different from those in the general field. It is necessary to propose targeted knowledge graph construction and reasoning techniques to support knowledge graphs in the medical field. Assistant intelligent applications in .

Contents of the invention

The purpose of the present invention is to solve the problem of automatic construction and reasoning of medical knowledge graphs.

For this reason, the present invention proposes a method for inferring medical field knowledge based on random walks, which mainly includes three parts:

(1) A named entity recognition method in the medical field based on contextual character binary groups and information entropy;

(2) A method for extracting relationships between entities in the medical field based on predicate sentiment classification;

(3) A reasoning method based on random walk knowledge map in medical field.

The specific content is as follows:

Use method (1) to identify named entities in the medical field, including concepts such as drugs, diseases, symptoms, populations, and ingredients; use method (2) to extract positive and negative relationships between named entities in the medical field, including applicable, taboo, etc. ;Using named entities in the medical field and the relationship between entities to automatically construct a medical knowledge graph, and using method (3) to achieve medical knowledge graph reasoning. Based on the above method, the automatic construction of medical knowledge map and knowledge reasoning in medical field are realized.

(1) A named entity recognition method in the medical field based on contextual character dyads and information entropy.

Collect conventional corpus and medical professional corpus, remove the punctuation marks and stop words, and establish two character transition probability matrices according to the medical corpus and the context in the conventional prediction database. Each element in the matrix is the transition frequency value in the context. Let Mat _medical be the context character transition probability matrix of the medical corpus, Mat _normal be the context character transition probability matrix of the conventional corpus, let {c _i , _ci+1 } be the continuous character context in the corpus, and calculate {c _i , c _i+1 }Transfer probability between medical corpus and regular corpus, we get matrix Mat _medical ( _ci ,ci ₊₁ ) and matrix Mat _normal ( _ci ,ci ₊₁ ).

Based on the context character transition probability matrix of the medical corpus and conventional corpus, information entropy is used to calculate the significance of each group of character contexts belonging to the medical field. Since the character transition probability in the conventional corpus is relatively stable, the medical corpus significantly deviates from the conventional corpus character transition probability. The character context determines the pharmaceutical named entity.

Calculate the information entropy of the character transfer probability according to the following formula, the information entropy Entropy(ci,ci ₊₁ ) is used to mark whether { _ci , _{ci+1 }} is a named entity in the medical field, if Entropy( _ci ,ci+ _{1 +1} )>t, where t(t=1) is a critical value, then {c _i , c _i+1 } is the character context of the same named entity, and continuous character contexts are combined to form a medical named entity.

(2) A method for extracting relationships between entities in the medical field based on predicate sentiment classification.

The medical corpus is segmented according to the punctuation marks to obtain a set of short sentences, and the emotions of some of the short sentences are marked, and the labels include positive, negative, and neutral. All short sentences with emotional tags in the medical corpus are processed by Chinese word segmentation using Viterbi-based conditional random field method, and all words are vectorized by word vector method. The word vectors of all words are weighted and averaged to obtain the text vectors of short sentences, and the support vector machine is used to train the text vectors with emotional labels to obtain the text emotion classification model. Based on this model, sentiment classification is carried out for all short sentences in the medical corpus, and short sentences with significant positive or negative emotions are extracted.

Perform Chinese word segmentation processing on the above short sentences with positive or negative emotions, and perform part-of-speech tagging on the words and phrases in them, and extract the predicates (verbs) in them. If the number of medical named entities contained in the short sentence is greater than or equal to 2, and the two sides of the predicate belong to the position or the predicate is the head word and the tail word, then extract the entities on both sides of the predicate and establish the relationship between the entities, and at the same time, according to the short sentence The relationship between positive or negative emotion discrimination entities belongs to positive relationship or negative relationship.

(3) A reasoning method based on random walk knowledge map in medical field.

According to the above-mentioned medical named entity recognition method and the relationship extraction method between entities, a medical knowledge graph KG(V, E, P) is constructed based on the triple representation method, where V represents the vertex in the knowledge graph, that is, the medical entity, and E represents The edge between two vertices in the knowledge graph is the relationship between two entities, and P represents the positive or negative attribute of the edge in the knowledge graph.

As shown in Figure 1, the relationship between knowledge map concepts, the concepts of medical entities include diseases, symptoms, medicines, groups of people, departments, body parts, etc. The vertices in the knowledge graph also include disease entities, such as colds, gastritis, etc.; symptom entities, such as cough, stomach pain, etc.; drug entities, such as aspirin, cephalosporin, etc.; crowd entities, such as babies, pregnant women, etc.; body part entities, such as head, Chest etc. The edge represents the relationship between each two entities, for example, cold-cough, gastritis-stomach pain means that cold and gastritis cause cough and stomach pain respectively. In addition, the relationship between medical entities also includes positive and negative relationships. For example, a cold-cough is a positive relationship, because a cold causes a cough, but a tetracycline-pregnant woman is a negative relationship, because tetracycline is a contraindication for pregnant women. medicine. Positive relationships include applying to, causing, etc., and negative relationships include cautious use, taboo, etc. For example, given a short sentence "Gastritis is inflammation of the gastric mucosa... usually manifested as epigastric pain, nausea, vomiting... Complications include bleeding, gastric ulcer...", the disease vertex extracted from the short sentence is {gastritis}, The symptom vertex extracted is {upper abdominal pain, nausea, vomiting... bleeding, gastric ulcer...}, and the relationship and weight are {(gastritis, upper abdominal pain, 1.0), (gastritis, nausea, 1.0), (gastritis, vomiting , 1.0)...(gastritis, bleeding, 1.0), (gastritis, gastric ulcer, 1.0)...}.

According to the medical knowledge map, knowledge reasoning is performed based on the random walk method. The reasoning process will be transformed into a traversal process, starting from limited clues (several entities) to iteratively search for reasoning results. V={v ₁ ,v ₂ ,...,v _n } is a group of entities, and the candidate results can be inferred from this group of entities according to the following formula.

Among them, score(v _i ) is the score of the specified entity v _i , In(v _i ) is the in-degree of v _i , Out(v _i ) is the out-degree of v _i , p _j,i are v _i and v _j The attribute value of the edge in between is 1 for the positive direction and -1 for the negative direction, and α (α=0.85) is an empirical parameter. During inference, the known group of entities is initialized on the knowledge graph, the corresponding vertex scores are initialized to 1, and the remaining vertex scores are initialized to 0. The scores of all vertices are obtained through random walk iterative calculation, and the scores are sorted. The entities corresponding to the vertices with higher scores are the candidate results that can be inferred from this group of known entities, and finally screened according to the actual situation.

Description of drawings

Figure 1 is a diagram of the relationship between knowledge map concepts

Detailed ways

The steps of the present invention are as follows:

Step 1: Collect regular corpus and medical professional corpus, and remove punctuation marks and stop words.

Step 2: According to the character context {c _i ,c _i+1 } in the medical corpus and the conventional prediction library, the character transition probability matrices Mat _medical ( _ci ,ci ₊₁ ) and Mat _normal ( c _i ,c _i+1 ).

Step 3: Based on the context character transition probability matrix of the medical corpus and the conventional corpus, the information entropy is used to calculate the significance of each group of character contexts belonging to the medical field.

Step 4: If information entropy Entropy(ci,ci ₊₁ )>t, where t(t=1) is the critical value, then { _ci , _{ci+1 }} is the character context of the same named entity, the combination Consecutive character contexts form pharmaceutical named entities.

Step 5: Segment the medical corpus according to punctuation marks to obtain a set of short sentences, and label the emotions of some of the short sentences, including positive, negative, and neutral.

Step 6: Use Chinese word segmentation and word vector method to vectorize all words. The word vectors of all words are weighted and averaged to obtain the text vectors of short sentences, and the support vector machine is used to train the text vectors with emotional labels to obtain the text emotion classification model. Sentiment classification of short sentences in medical corpus is carried out based on this model.

Step 7: Use the part-of-speech tagging method to extract the predicates in the sentence. If the number of medical named entities contained in the sentence is greater than or equal to 2, and belong to both sides of the predicate or the predicate is the head word and tail word, then extract both sides of the predicate Entities and establish the relationship between entities, and at the same time judge whether the relationship between entities is a positive relationship or a negative relationship according to the positive or negative emotion of the short sentence.

Step 8: According to the medical named entity recognition and the relationship between entities, based on the triple representation, the medical knowledge graph is constructed.

Step 9: According to the medical knowledge map, knowledge reasoning is performed based on the random walk method. The reasoning process will be transformed into a traversal process, starting from limited clues (several entities) to iteratively search for candidate reasoning results. .

Claims (4)

1. A medicine field knowledge inference method based on random walk is characterized by comprising the following steps:

(1) a medicine field named entity identification method based on context character binary and information entropy;

(2) a method for extracting relationships between medical field entities based on predicate sentiment classification;

(3) a medicine field knowledge graph reasoning method based on random walk.

2. The method for identifying a named entity in the medical field based on a context character binary group and information entropy as claimed in claim 1, wherein the named entity in the medical field is identified by comparing the statistical representation of the character context of the named entity in the general field with the statistical representation of the character context of the named entity in the medical field by using a context character binary group and information entropy method, aiming at the problem that the traditional named entity identification method is inaccurate due to the fact that the statistical representation of the character context of the named entity in the medical field is not smooth.

3. The method for extracting relationships between medical field entities based on predicate sentiment classification as claimed in claim 1, wherein the relationships between medical field entities are extracted by carrying out sentiment classification on adjacent predicates aiming at positive and negative relationships between medical field entities and related to predicate sentiment between entities.

4. The random walk based reasoning method for knowledge base of medical field according to claim 1, wherein the random walk method is used to perform reasoning for medical knowledge in the knowledge base of medical science, in order to solve the problem that the knowledge base is difficult to be used directly for reasoning due to the intensive and complicated incidence relation between the entities of the knowledge base of medical field.