CN104750819A - Biomedicine literature search method and system based on word grading sorting algorithm - Google Patents

Abstract

A biomedical literature retrieval method and system based on a word grouping sorting algorithm, the retrieval method includes S1, a search engine query extraction step; S2, a candidate extended vocabulary extraction step; S3, a candidate extended vocabulary feature extraction and labeling step; S4, a candidate The step of training the extended vocabulary sorting model; S5, the online search engine query and extraction step; S6, the online candidate extended vocabulary extraction and its feature extraction and scoring steps; S7, the query result return step. The retrieval system includes a search engine query extraction module, a candidate extended vocabulary extraction module, a candidate extended vocabulary feature extraction and labeling module, a candidate extended vocabulary sorting model training module, a query reconstruction module, and a query result return module. From the perspective of query expansion, the invention uses word grouping and sorting algorithms and inherent dictionary resources in the biomedical field to select professional vocabulary that best expresses user information needs, completes retrieval tasks, and improves retrieval performance.

Description

A Biomedical Literature Retrieval Method and System Based on Word Grouping Algorithm

technical field

The invention relates to the technical fields of data mining and search engines, in particular to a biomedical document retrieval method and system based on word grouping and sorting algorithms.

Background technique

In recent years, with the rapid development of the field of biomedicine (Biomedicine), biomedical research has achieved many valuable results, which not only contributed to the treatment of some diseases that once seemed difficult to solve, but also from a more far-reaching perspective. , It also promotes the development and deepening of human beings' understanding of themselves.

However, with the rapid increase in the number of biomedical literature, the amount of relevant information is also increasing exponentially. The massive amount of literature and information has brought difficulties for biomedical researchers and related practitioners to obtain information. Traditional manual information acquisition The methods have gradually become inapplicable. Therefore, it is necessary to use information retrieval technologies and methods to assist relevant personnel to obtain the required information.

Traditional information retrieval technology can sort documents or web pages according to the query submitted by users, and return the sorting results to users. However, it is difficult to obtain better retrieval performance by directly applying traditional information retrieval methods to the retrieval tasks of biomedical literature. The reason is that the inherent characteristics of the biomedical field are not fully considered. For example, the biomedical field has many professional vocabularies, and these professional vocabularies often have many synonyms and abbreviations at the same time. If the characteristics of the biomedical field can be fully considered in the traditional information retrieval methods, the performance of biomedical information retrieval will be further improved.

Query expansion technology is one of the key technologies in the traditional information retrieval field. It can supplement and improve the query based on the original query submitted by the user and according to the user's retrieval intention, so as to obtain a query that is more in line with the user's retrieval intention and improve the retrieval performance. The existing query expansion methods can be divided into two categories: one is the query expansion method based on the document collection, this kind of method takes the whole data document collection or part of the data document collection as the research object, extracts the content related to the query from it, and perfects the query expansion method. Original query; the other is query expansion technology based on external extended resources. External resources mainly include dictionary resources, retrieval system query logs, anchor text and Wikipedia, etc. Many studies have shown that using external extended resources to improve original queries can be better. Complete the query expansion task, thereby improving the retrieval performance.

Since there are many resources in the field of biomedicine, such as dictionaries, if these resources can be fully utilized in the process of information retrieval to supplement and improve the queries submitted by users, the retrieval performance will be greatly improved.

To establish a literature search for the biomedical field, one should first understand the characteristics and resources of this field. There are a large number of professional vocabularies in the literature in the field of biomedicine, and these vocabularies contain many synonyms and abbreviations and other complex situations, which brings great challenges to the establishment of retrieval systems. For example, for the drug paracetamol, its English The name is paracetamol, and in the International Standard Classification of Drugs, its name is acetaminophen (acetaminophen), and its scientific name in the field of medicinal chemistry is C8H9NO2 or NO2BE01. For the above multiple names, if only It is difficult to retrieve all relevant documents by querying one of the names. Fortunately, there are still many inherent knowledge bases and resources in the biomedical field, such as Medical Subject Headings (MeSH: Medical Subject Headings) and Gene Ontology (GO: Gene Ontology). Utilizing these resources will greatly improve the performance of biomedical literature retrieval.

The learning to rank algorithm is a general term for a series of supervised learning algorithms used to sort documents in information retrieval. Its main feature is the application of machine learning technology to solve the ranking problem in information retrieval, and has achieved better results. retrieval and sorting performance. The ranking problem can also be regarded as a selection problem of the optimal item. Therefore, in recent years, the ranking learning algorithm has been applied to many other tasks, such as recommending corresponding items for users in recommendation systems based on the historical information of users and items. wait.

Contents of the invention

The purpose of the present invention is to provide a biomedical literature retrieval method based on the word grouping sorting algorithm that can provide users with more accurate biomedical literature, more effectively meet the information needs of users, and effectively supplement and improve user queries. system.

The technical solution adopted by the present invention to solve the problems of the prior art: a biomedical literature retrieval method based on word grouping and sorting algorithm, including the following offline training phase and online query phase, wherein the offline training phase includes the following steps:

S1, search engine query extraction step: according to the historical query records of the search engine, extract multiple groups of queries and the first N query result documents obtained in each query; and collect the queries and query result documents into a query pool, where N is a natural number;

S2. Candidate extended vocabulary extraction step: extract the professional vocabulary in the first N query result documents of each query in the query pool according to the biomedical resources, and obtain the number of occurrences of each professional vocabulary in the query result document by counting Or the weighted sum of the number of occurrences; according to the weighted and descending order of the number of times or times that each professional word appears in the query result document, select the M professional words with the highest number of occurrences or the weighted sum of the highest number of times as candidate extended words, where M is a natural number;

S3. Feature extraction and labeling steps of candidate extended vocabulary:

The feature extraction and labeling of the candidate extended vocabulary are carried out at the same time; among them, the relevance labeling of the candidate expanded vocabulary is marked by comparing the retrieval performance of the original query with the retrieval performance when the candidate expanded vocabulary is added to the original query; the retrieval performance High and low evaluation indicators include: accuracy rate, average accuracy rate, NDCG value and MRR value; the specific method of correlation labeling is as follows:

$\mathrm{<span\; class="notranslate">\; label</span>}<span\; class="notranslate">\; =</span>\left\{\begin{array}{cc}\mathrm{<span\; class="notranslate">\; 1</span>}& \mathrm{<span\; class="notranslate">\; eval</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; query</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; term</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ></span>\mathrm{<span\; class="notranslate">\; eval</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; query</span>}<span\; class="notranslate">\; )</span>\\ \mathrm{<span\; class="notranslate">\; 0</span>}& \mathrm{<span\; class="notranslate">\; eval</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; query</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; term</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&le;</span>\mathrm{<span\; class="notranslate">\; eval</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; query</span>}<span\; class="notranslate">\; )</span>\end{array}\right\}$

Among them, eval() is the evaluation index function used to evaluate the retrieval performance, eval(query+term) is the score of the evaluation index function eval() when adding the candidate extended vocabulary term to the query query, and eval(query) is The score of the evaluation index function when evaluating the query query; the label marked as 1 indicates that the candidate extended vocabulary is related to the query query; the label marked as 0 indicates that the candidate expanded vocabulary is not related to the query query;

The feature extraction of candidate extended vocabulary is to extract the distribution information of candidate extended vocabulary, the distribution information of candidate vocabulary in biomedical resources, and the candidate extended vocabulary from the biomedical resources and the first N query result documents returned by the query in the query pool. and the correlation information of the original query to prepare for the training ranking model, and after extracting multiple features of the same candidate extended vocabulary, normalize all feature values to control all feature values in [0,1] On the interval, the normalization process is as follows:

$\mathrm{<span\; class="notranslate">\; newFeatureValue</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; oldFeatureValue</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; min</span>}\mathrm{<span\; class="notranslate">\; value</span>}}{\mathrm{<span\; class="notranslate">\; max</span>}\mathrm{<span\; class="notranslate">\; value</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; min</span>}\mathrm{<span\; class="notranslate">\; value</span>}}$

Among them, minValue and maxValue are the minimum and maximum values of a feature, respectively;

S4. Training steps of the candidate extended vocabulary sorting model: according to the relevance degree labeling and various features of the candidate expanded vocabulary, use the word grouping sorting algorithm to train and obtain the weight value of each feature. The specific steps are: select one that is marked as relevant in step S3 Candidate extended vocabulary and a number of candidate extended vocabulary marked as irrelevant form a word group, and select several such word groups as training samples; randomly assign initial weights to the features of each candidate extended word, and use the feature weighted score to evaluate each According to the ranking results of each word group, the overall ranking loss is calculated, and the weight of each dimension feature is dynamically adjusted according to the gradient value of the loss function, where the ranking loss is: Among them, NumSample is the number of candidate extended vocabulary groups in the word grouping, loss _i is the loss value of each word grouping, the loss value is obtained by calculating the sorting position of the related expanded vocabulary, and the higher the ranking position is, the smaller the corresponding loss value is; by Iterate the previous process in a loop until the overall loss value is less than a certain threshold or reaches the specified number of iterations and the training is completed, and the final selected feature value is used as the sorting model for the training completion;

The online inquiry phase includes the following steps:

S5. Online search engine query and extraction steps: For the new query submitted by the user online, retrieve the first N1 query results; extract the professional vocabulary and its various features in the first N1 search results according to biomedical resources, where N1 is a natural number;

S6. Online Candidate Extended Vocabulary Extraction and Feature Extraction and Scoring Steps: According to the biomedical resources, use the candidate extended vocabulary extraction method in the offline stage S2-S3 and the feature extraction method for the candidate extended vocabulary to search for the first N1 search results In the online query stage, the professional vocabulary and its various features are extracted to obtain the candidate extended vocabulary in the online query stage, and the extracted features are used to measure the importance of the candidate extended vocabulary in the extended query; according to the feature weight obtained in step S4 training, the online query Stage candidate extended vocabulary is scored, and K1 candidate extended vocabulary with the highest score is selected to be added to the new query submitted online as an extended query, where K1 is a natural number;

For a candidate extended vocabulary in the online query stage that is annotated and extracted using biomedical resources, its score is Where FeatureNum is the total number of features, ai is the weight value of the i-th feature in the ranking model, and feature _i (term) is the feature value of the i-th feature corresponding to the candidate extended vocabulary term in the online query phase;

Rank them according to the scores of the candidate extended vocabulary in the online query stage, and select the top K1 candidate extended vocabulary in the online query stage as the extended vocabulary to add to the new query submitted online, the added candidate extended vocabulary in the online query stage is in The weights in the extended query can be expressed as $\mathrm{weight}={\mathrm{\&Sigma;}}_{i=1}^{\mathrm{count}}{\mathrm{weight}}_i\&CenterDot;{\mathrm{feature}}_i+\mathrm{sign}\&Center\; Dot;{\mathrm{weight}}_{\mathrm{original}},$ Wherein sign is a sign function, sign=1 when the candidate extended vocabulary in the online query stage appears in the new query submitted online, otherwise sign=0, weight _original is the weight value of the new query submitted online in the extended query;

S7. Step of returning query results: performing retrieval according to the extended query, and returning the retrieval results to the user.

In step S2, the weighted sum of the number of occurrences of professional words in the query result document is Among them, count _i is the number of times the word appears in the i-th document, and d(i) is the decay factor of the i-th document.

In step S3, the evaluation index function eval() is the average accuracy rate function, namely:

${\mathrm{<span\; class="notranslate">\; eval</span>}}_{\mathrm{<span\; class="notranslate">\; MAP</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{{\mathrm{<span\; class="notranslate">\; RelDoc</span>}}_{\mathrm{<span\; class="notranslate">\; query</span>}}}<span\; class="notranslate">\; \&CenterDot;</span>{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{{\mathrm{<span\; class="notranslate">\; RelDoc</span>}}_{\mathrm{<span\; class="notranslate">\; query</span>}}}\frac{\mathrm{<span\; class="notranslate">\; i</span>}}{\mathrm{<span\; class="notranslate">\; rank</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span>}$

Wherein, RelDoc _query is the number of related documents of a given query query, and rank(i) indicates the position of the i-th related document in the ranking list of document results.

In step S1, when there is no historical query record, manually obtain the records of the query and its results by constructing a biomedical query and retrieval method; the retrieval method adopts the vector space model, BM25 retrieval model or based on different smoothing The language model for the method.

The loss value in step S4 is: Wherein, rank _i is the ranking position of the relevant candidate expansion words in the word grouping list.

Biomedical resources refer to dictionaries or knowledge bases containing biomedical vocabulary.

The characteristics of the candidate extended vocabulary include the frequency TF of the candidate extended vocabulary in the result document, the TF-IDF value of the candidate extended vocabulary, the number of documents where the candidate extended vocabulary and the original query co-occur, the candidate extended vocabulary and the original query in the same The number of co-occurrences in the text window, the number of occurrences of candidate extended vocabulary in biomedical resources, the number of term concepts containing the candidate extended vocabulary in biomedical resources, and the inclusion relationship between biomedical term concepts.

A biomedical literature retrieval system based on the word grouping sorting algorithm, comprising an offline training part and an online retrieval part; the offline training part includes the following parts:

Search engine query extraction module: used to extract multiple sets of queries and the first N query result documents obtained in each query according to the historical query records of the search engine; and collect the queries and query result documents into a query pool, where N is a natural number;

Candidate Extended Vocabulary Extraction Module: It is used to extract professional vocabulary from the first N query result documents obtained by the search engine query extraction module by using the inherent resources in the biomedical field when a user query is given, and the professional vocabulary in the The frequency of occurrence in the query result document or the weighted sum of the number of occurrences is recorded; according to the number of occurrences of each professional word in the query result document or the weighted and descending order of the number of occurrences, select the M professional words with the highest frequency of occurrence as candidate extensions Vocabulary, where M is a natural number;

Feature extraction and labeling module of candidate extended vocabulary: used to extract features related to the candidate extended vocabulary obtained in the candidate extended word extraction module, and mark the correlation of candidate extended vocabulary according to the impact of candidate extended vocabulary on retrieval performance degree;

Candidate extended vocabulary sorting model training module: used to use the word grouping sorting algorithm to train the vocabulary sorting model to obtain the weight value of each feature of the candidate extended vocabulary after extracting the features of the candidate extended vocabulary and marking the degree of relevance of the candidate extended vocabulary;

The online search section includes:

Query reconstruction module: used for professional vocabulary extraction and candidate extended vocabulary scoring in new queries; including online search engine query extraction module, online candidate extended vocabulary extraction and feature extraction and scoring modules, in which the online search engine query extraction module is used Based on the new query submitted by the user online, the first N1 query results are retrieved; the professional vocabulary and its various features in the first N1 search results are extracted according to biomedical resources, where N1 is a natural number. The online candidate extended vocabulary extraction and feature extraction and scoring module uses the weight value of the candidate extended vocabulary output by the vocabulary ranking model to calculate the corresponding weight, and adds it to the original query to obtain the extended query;

Query result return module: used to return the result document retrieved by the extended query to the user.

The beneficial effect of the present invention is that: the present invention mainly proceeds from the perspective of query expansion, and selects the professional vocabulary that can best express the user's information needs by using word grouping sorting algorithms and inherent dictionaries in the biomedical field in the query expansion, which is more effective The task of retrieval can be completed in order to provide users with more appropriate retrieval results. The present invention uses resources in the field of biomedicine to supplement and improve the original query, thereby improving the performance of retrieval. When using the TREC gene task literature data set as the data set and using the traditional BM25 retrieval model as the benchmark retrieval model for literature retrieval, a literature retrieval accuracy rate of 25.62% can be obtained; and on this basis, the method and system involved in the present invention are used When searching, the literature retrieval accuracy rate of 26.30% can be obtained, the retrieval performance has been significantly improved, and the literature retrieval system involved in the present invention can effectively retrieve the biomedical literature most relevant to the user's query, improving user satisfaction.

Description of drawings

Fig. 1 is a schematic flow chart of the retrieval method of the present invention;

Fig. 2 is a schematic diagram of the logical structure of the retrieval system of the present invention.

Detailed ways

The present invention is described below in conjunction with accompanying drawing and specific embodiment:

Fig. 1 is a schematic flow chart of a biomedical literature retrieval method based on word grouping and sorting algorithm of the present invention, a kind of biomedical literature retrieval method based on word grouping and sorting algorithm, comprises following offline training stage and online query stage, wherein, offline training Phase consists of the following steps:

S1, search engine query extraction step: according to the historical query records of the search engine, extract multiple groups of queries and the first N query result documents obtained in each query; and collect the queries and query result documents into a query pool, where N is Natural number. In this embodiment, N=10;

Among them, the historical query records of the search engine mainly refer to the query history recorded by the biomedical literature retrieval system and the corresponding query results. These queries and the corresponding query results will be used for training the ranking model in the offline state.

When there is no relevant historical query record, the records of query and retrieval results can be obtained manually by constructing biomedical query and retrieval. The retrieval method can adopt various sorting models in traditional information retrieval, including but not limited to vector space model, BM25 retrieval model, language model based on different smoothing methods, etc.

S2. Candidate extended vocabulary extraction step: extract the professional vocabulary in the first N query result documents of each query in the query pool according to the biomedical resources, and obtain the number of occurrences of each professional vocabulary in the query result document by counting Or the weighted sum of the number of occurrences; according to the number of times each professional word appears in the query result document or the weighted and descending order of the number of times, select the M professional words with the highest frequency of occurrence or the highest weighted sum of times as candidate extended words, where M is Natural number;

Among them, biomedical resources refer to resources such as dictionaries or knowledge bases containing biomedical professional vocabulary, including but not limited to: Medical Thesaurus (MeSH), Gene Ontology (GO) and the super vocabulary released by the Unified Medical Language System (UMLS) Library (Metathesaurus), Semantic Network (Semantic Network) and expert semantic dictionary tools (SPECIALIST Lexicon and Lexical Tools), etc.

Taking the medical thesaurus MeSH as the biomedical resource used in the present invention as an example, the professional vocabulary in the first N query result documents corresponding to the query is extracted, and each professional vocabulary extracted corresponds to its occurrence in the document The number of times or the weighted sum of the number of occurrences. For example, the professional vocabulary term is weighted by the number of times it appears in the first N documents Calculated, where count _i is the number of times the word appears in the i-th document, d(i) is the attenuation factor of the i-th document, and the weighted sum of the number of professional words is used to weight the frequency of words that appear in different documents. Therefore, the word frequency in the documents ranked higher has a greater weight, and the control makes the professional words contained in the documents ranked lower obtain less scores. According to the above formula Sort the selected professional vocabulary from high to low in the value of count(term), or sort the selected professional vocabulary according to the value of score(term) from high to low, and select the top M with the highest ranking vocabulary as a candidate extended vocabulary, and the value of M is 150 in this embodiment.

S3. Steps of feature extraction and correlation labeling of candidate extended vocabulary:

The feature extraction and labeling of the candidate extended vocabulary are carried out at the same time; wherein, the relevance labeling of the candidate extended vocabulary is realized by comparing the retrieval performance of the original query with the retrieval performance when the extended vocabulary is added to the original query. The idea of tagging the candidate extended vocabulary is: add a single candidate extended vocabulary to the original query for retrieval, and if the performance of the retrieval results is improved, mark the extended vocabulary as relevant to the original query. The evaluation indicators of retrieval performance include but are not limited to: accuracy rate (Precision), average accuracy rate (MAP), NDCG value and MRR value, etc. The specific way of labeling is as follows:

$\mathrm{<span\; class="notranslate">\; label</span>}<span\; class="notranslate">\; =</span>\left\{\begin{array}{cc}\mathrm{<span\; class="notranslate">\; 1</span>}& \mathrm{<span\; class="notranslate">\; eval</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; query</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; term</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ></span>\mathrm{<span\; class="notranslate">\; eval</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; query</span>}<span\; class="notranslate">\; )</span>\\ \mathrm{<span\; class="notranslate">\; 0</span>}& \mathrm{<span\; class="notranslate">\; eval</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; query</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; term</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&le;</span>\mathrm{<span\; class="notranslate">\; eval</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; query</span>}<span\; class="notranslate">\; )</span>\end{array}\right\}$

Among them, eval() is the evaluation index function used to evaluate the retrieval performance, eval(query+term) is the score of the evaluation index function eval() when adding the candidate extended vocabulary term to the given query query, eval(query+term) ) is the score of the evaluation index function when evaluating a given query query. When the evaluation score of retrieval with the original query plus a candidate vocabulary is greater than the evaluation score of the original query itself, the candidate extended vocabulary is marked as 1, which means that the vocabulary is related to the original query; and When the evaluation score of the original query plus a candidate vocabulary is not greater than the evaluation score of the original query itself, the candidate extended vocabulary is marked as 0, which means that the vocabulary is irrelevant to the original query.

In this embodiment, the evaluation function eval () is the average accuracy rate, namely:

${\mathrm{<span\; class="notranslate">\; eval</span>}}_{\mathrm{<span\; class="notranslate">\; MAP</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{{\mathrm{<span\; class="notranslate">\; RelDoc</span>}}_{\mathrm{<span\; class="notranslate">\; query</span>}}}<span\; class="notranslate">\; \&CenterDot;</span>{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{{\mathrm{<span\; class="notranslate">\; RelDoc</span>}}_{\mathrm{<span\; class="notranslate">\; query</span>}}}\frac{\mathrm{<span\; class="notranslate">\; i</span>}}{\mathrm{<span\; class="notranslate">\; rank</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span>}$

Among them, RelDoc _query is the number of relevant documents of a given query query, and rank(i) indicates the position of the i-th related document in the document result ranking list, for example, rank(3)=5 indicates that it is in the result ranking list The 3rd related document appears in the 5th position of the sorted list.

The feature extraction of candidate extended vocabulary is to extract the distribution information of candidate extended vocabulary, the distribution information of candidate vocabulary in biomedical resources, and the candidate extension The correlation information of the vocabulary and the original query is used to prepare for the training ranking model, and after extracting multiple features of the same candidate extended vocabulary, normalize all feature values; to control all feature values in [0, 1 ] interval, the specific process of normalization is:

$\mathrm{newFeatureValue}=\frac{\mathrm{oldFeatureValue}-\min \mathrm{value}}{\max \mathrm{value}-\min \mathrm{value}},$ minValue and maxValue are the minimum and maximum values of a feature, respectively.

Among them, the characteristics of extended vocabulary include:

1. The frequency TF of the candidate extended vocabulary appearing in the result document. This feature can be obtained according to the number of occurrences of the professional vocabulary term in the result document.

2. The TF-IDF value of the candidate extended vocabulary. TF-IDF is one of the classic models in the field of information retrieval, which can be used to measure the relative importance of words. The calculation method is shown in the following formula:

${\mathrm{<span\; class="notranslate">\; score</span>}}_{\mathrm{<span\; class="notranslate">\; TF</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; IDF</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; count</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; term</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&Center\; Dot;</span>\mathrm{<span\; class="notranslate">\; log</span>}\frac{\mathrm{<span\; class="notranslate">\; TotalDoc</span>}}{\mathrm{<span\; class="notranslate">\; df</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; term</span>}<span\; class="notranslate">\; )</span>}$

Among them, count(term) is the number of times the candidate extended vocabulary appears in the result document i, TotalDoc is the total number of documents in the training data, and df(term) is the number of documents in which the candidate extended vocabulary appears.

3. The number of documents in which the candidate extended vocabulary co-occurs with the original query. This feature can be used to calculate how relevant the original query is to the candidate expanded vocabulary.

4. The number of times the candidate expanded vocabulary co-occurs with the original query in the same text window. This feature is used to calculate the degree of correlation between the query word in the original query and the candidate extended vocabulary within a certain range, where the text window refers to the range of documents in which the original query word and the candidate word appear in the same document, and the extended vocabulary is related to the original The number of words that are spaced between query words.

5. The number of occurrences of candidate expanded vocabulary in biomedical resources such as MeSH. This feature is used to calculate and measure the part information of the candidate extended vocabulary in biomedical resources.

6. In biomedical resources such as MeSH, the number of term concepts that include the candidate expanded vocabulary. There is often an inclusion relationship between the concepts of biomedical terminology, and this feature can also measure the importance of a candidate word in biomedical resources.

Among the features of candidate extended vocabulary extracted above, feature 1 and feature 2 are used to measure the distribution information of candidate extended vocabulary in the document collection; feature 3 and feature 4 are used to measure the degree of relevance between candidate extended vocabulary and the original query; and feature 5 and feature 6 are used to measure the distribution information of candidate extended vocabulary in biomedical resources. The extended vocabulary features involved in the present invention include but are not limited to the above-mentioned features. Through the extraction of the above-mentioned multiple features, they can be used as the input of word grouping and sorting algorithms to better measure the importance of candidate extended vocabulary.

S4. Training step of the candidate extended vocabulary sorting model: according to the relevant degree labeling and various features of the candidate expanded vocabulary obtained in step S3 as input, use the sorting model training of the word grouping sorting algorithm to obtain the weight value of each feature. The specific steps are as follows: Select a candidate extended vocabulary that is marked as relevant in step S3 (that is, the corresponding candidate extended vocabulary when the label is 1) and several candidate extended vocabulary that are marked as irrelevant (that is, the corresponding candidate extended vocabulary when the label is 0) Form a word group, select several such word groups as training samples; randomly assign initial weights to the word features of each candidate extended vocabulary, and sort the related extended vocabulary in each word group through the feature weighted score; according to each word For the sorting results of the group, calculate the overall sorting loss, and dynamically adjust the weight of each dimension feature according to the gradient value of the loss function, where the sorting loss is: Among them, NumSample is the number of candidate extended vocabulary groups in the word grouping, and loss _i is the loss value of each word grouping. The loss value is obtained by calculating the sorting position of the related expanded vocabulary. The higher the sorting position, the smaller the corresponding loss value; Iterate the previous process in a loop until the overall loss value is less than a certain threshold or reaches the specified number of iterations, and the training is completed, and the finally selected feature value is used as the sorting model for the training completion; in this embodiment, 100 iterations are selected to terminate the training.

In this example, the loss value is: Among them, rank _i is the ranking position of the related candidate expansion words in the word grouping list. When it is ranked first, the loss is 0, and when it is ranked last, the loss is maximized. In addition, the calculation formula of the loss value includes but is not limited to this calculation formula.

In the ranking model, the formula for calculating the final score of the extended vocabulary is as follows:

$\mathrm{<span\; class="notranslate">\; score</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; term</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; FeatureNum</span>}}{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; \&CenterDot;</span>{\mathrm{<span\; class="notranslate">\; feature</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; term</span>}<span\; class="notranslate">\; )</span>$

Among them, FeatureNum is the total number of features, a _i is the weight value of the i-th feature, and feature _i (term) is the feature value of the i-th feature corresponding to the candidate word term. The ranking model trained here can be used to test the selection of query-related extended vocabulary. The above steps are completed offline.

The online inquiry phase includes the following steps:

S5. Online search engine query and extraction steps: For the new query submitted by the user online, retrieve the first N1 query results; extract the professional vocabulary and its various features in the first N1 search results according to biomedical resources, where N1 is a natural number;

It should be noted that this step refers to the online situation. After the user submits the query to the biomedical literature search engine, this method will automatically obtain the N1 query results with the highest ranking in the initial search, which is used to expand the user query. And so on, the processing is transparent to the user.

S6. Online Candidate Extended Vocabulary Extraction and Feature Extraction and Scoring Steps: According to the biomedical resources, use the candidate extended vocabulary extraction method in the offline stage S2-S3 and the feature extraction method for the candidate extended vocabulary to search for the first N1 search results In the online query stage, the professional vocabulary and its various features are extracted to obtain the candidate extended vocabulary in the online query stage, and the extracted features are used to measure the importance of the candidate extended vocabulary in the extended query; according to the feature weight obtained in step S4 training, the online query Stage candidate extended vocabulary is scored, a new query is constructed according to the score, and K1 online query stage candidate extended vocabulary with the highest score is selected to be added to the new query submitted online as the online stage extended query, where K1 is a natural number;

For a candidate extended vocabulary in the online query stage that is annotated and extracted using biomedical resources, its score is Where FeatureNum is the total number of features, a _i is the weight value of the i-th feature in the ranking model, and feature _i (term) is the feature value of the i-th feature corresponding to the candidate extended vocabulary term in the online query phase;

Rank them according to the scores of online query candidate extended vocabulary, and select the top K1 words as the online stage candidate extended vocabulary to add to the new query, the weight of the added online stage candidate extended vocabulary in the extended query can be expressed as for $\mathrm{weight}={\mathrm{\&Sigma;}}_{i=1}^{\mathrm{count}}{\mathrm{weight}}_i\&CenterDot;{\mathrm{feature}}_i+\mathrm{sign}\&Center\; Dot;{\mathrm{weight}}_{\mathrm{original}},$ Wherein sign is a sign function, sign=1 when the candidate extended vocabulary of this online stage appears in the new query submitted online, otherwise sign=0, weight _original is the weight value of the new query submitted online in the extended query;

The specific form of the final extended query is as follows:

(weight ₁ query _original weight ₂ (w ₁ term ₁ w ₂ term ₂ … w _k term _k ))

Among them, weight ₁ is the weight of the new query submitted online in the extended query, weight ₂ is the weight of all newly added extended vocabulary in the extended query, w ₁ ,w ₂ ,...,w _K are the extended vocabulary term ₁ ,term ₂ ,...,term _K corresponds to the score weight, and K is the number of extended vocabulary finally selected. In this embodiment, the value of weight ₁ is 0.5, the value of weight ₂ is 0.5, and the value of K is 50.

S7. Step of returning query results: performing retrieval according to the extended query, returning the retrieval results to the user, and completing the retrieval process.

Corresponding to the above method, the present invention also provides a biomedical document retrieval system based on word grouping and sorting algorithm. Accompanying drawing 2 has provided the logical structural diagram of this system.

A biomedical literature retrieval system based on the word grouping sorting algorithm, comprising an offline training part and an online retrieval part; the offline training part includes the following parts:

Search engine query extraction module: used to extract multiple sets of queries and the first N query result documents obtained in each query according to the historical query records of the search engine; and collect the queries and query result documents into a query pool, where N is a natural number; the search engine query extraction module can retrieve the biomedical literature associated with the user query according to the user query, and return the search results to the user, while the operations and operations such as query expansion in the system are very important to the user It is transparent and cannot be seen.

Candidate Extended Vocabulary Extraction Module: It is used to extract professional vocabulary from the first N query result documents obtained by the search engine query extraction module by using the inherent resources in the biomedical field when a user query is given, and the professional vocabulary in the The number of times (frequency) or the weighted sum of the number of occurrences in the query result document is recorded; according to the number of times each professional word appears in the query result document or the weighted and descending order of the number of occurrences, select the M professional words with the highest number of occurrences As a candidate extended vocabulary, where M is a natural number;

Feature extraction and labeling module of candidate extended vocabulary: used to extract features related to the candidate extended vocabulary obtained in the candidate extended word extraction module, and mark the correlation of candidate extended vocabulary according to the impact of candidate extended vocabulary on retrieval performance degree; during offline training, the relevant degree labels and various features of the candidate extended vocabulary will be used as the input of the word grouping and sorting algorithm; during online query, this module is used to extract the feature information associated with the candidate extended vocabulary.

Candidate extended vocabulary sorting model training module: used to use the word grouping sorting algorithm to train the vocabulary sorting model to output the weight value of each feature of the candidate extended vocabulary after extracting the candidate extended vocabulary features and marking the degree of relevance of the candidate extended vocabulary; the weight value A measure of the importance of an expanded vocabulary that can be used for unknown queries.

The online search section includes:

Query reconstruction module: used for professional vocabulary extraction and candidate extended vocabulary scoring in new queries; including online search engine query extraction module, online candidate extended vocabulary extraction and feature extraction and scoring modules, in which the online search engine query extraction module is used Based on the new query submitted by the user online, the first N1 query results are retrieved; the professional vocabulary and its various features in the first N1 search results are extracted according to biomedical resources, where N1 is a natural number. The online candidate extended vocabulary extraction and its feature extraction and scoring module use the weights of candidate extended vocabulary output by the vocabulary ranking model to calculate the corresponding weights, and add them to the original query to obtain the extended query.

The query result returning module is used to return the result document retrieved by the extended query to the user. The returned result obtained by the user is actually the result of the returned result submitted by the user after query expansion, and the process of query expansion is invisible to the user.

Based on the above description of specific implementations of the method and system involved in the present invention, description will be made in conjunction with specific embodiments. In this embodiment, it is assumed that the user has completed the training of the sorting model through historical data. When the user submits a new query "mad cow disease", the system first retrieves the frequency information in the pre-examination document according to the word, Select the candidate extended vocabulary, and the top 10 extended vocabulary among the candidate extended vocabulary and their correlation annotations are shown in the following table:

ranking vocabulary Correlation 1 disease relevant 2 prions (prions) relevant 3 cause (cause) irrelevant 4 infectious relevant 5 conversion (conversion) irrelevant 6 cow (cattle) relevant 7 spongiform (sponge tissue) relevant 8 fatal (fatal) irrelevant 9 encephalopathies (epileptic encephalopathy) relevant 10 mad (crazy) relevant

It can be seen from the above table that among the top 10 candidate extended words, there are 3 irrelevant words, if they are directly added to the original query, it will have a negative impact on the retrieval performance. Next, the features related to the candidate extended vocabulary are extracted from the document and the biomedical dictionary MeSH, and the weight of each feature is obtained by using the ranking model, and all the candidate extended vocabulary are re-scored and sorted.

The top 10 extended vocabulary selected after sorting is shown in the table below. It can be seen from the table that the top 10 queries in the extended queries after sorting are all related words. Adding these queries to the original query according to their normalized ranking scores as weights for retrieval can further improve retrieval performance.

The description of the above embodiments explains and illustrates the biomedical literature retrieval method and system based on the word grouping and sorting algorithm provided by the present invention. The method and system can expand the original query submitted by the user by using resources such as the knowledge base in the biomedical field. In the expansion, the word grouping and sorting algorithm is used to expand the vocabulary importance measure, and the query submitted by the user is processed through the query expansion process. In order to supplement and improve, it ensures the accuracy of the query results and further satisfies the information needs of users.

The above content is a further detailed description of the present invention in combination with specific preferred technical solutions, and it cannot be assumed that the specific implementation of the present invention is limited to these descriptions. For those of ordinary skill in the technical field of the present invention, without departing from the concept of the present invention, some simple deduction or replacement can be made, which should be regarded as belonging to the protection scope of the present invention.

Claims (8)

1. A biomedical literature retrieval method based on word grouping sorting algorithm, is characterized in that, comprises following off-line training phase and online query phase, wherein, off-line training phase comprises the following steps: S1, search engine query extraction step: according to the historical query records of the search engine, extract multiple groups of queries and the first N query result documents obtained in each query; and collect the queries and query result documents into a query pool, where N is a natural number; S2. Candidate extended vocabulary extraction step: extract the professional vocabulary in the first N query result documents of each query in the query pool according to the biomedical resources, and obtain the number of occurrences of each professional vocabulary in the query result document by counting Or the weighted sum of the number of occurrences; according to the weighted and descending order of the number of times or times that each professional word appears in the query result document, select the M professional words with the highest number of occurrences or the weighted sum of the highest number of times as candidate extended words, where M is a natural number; S3. Feature extraction and labeling steps of candidate extended vocabulary: The feature extraction and labeling of the candidate extended vocabulary are carried out at the same time; among them, the relevance labeling of the candidate expanded vocabulary is marked by comparing the retrieval performance of the original query with the retrieval performance when the candidate expanded vocabulary is added to the original query; the retrieval performance High and low evaluation indicators include: accuracy rate, average accuracy rate, NDCG value and MRR value; the specific method of correlation labeling is as follows: $\mathrm{<span\; class="notranslate">\; label</span>}<span\; class="notranslate">\; =</span>\left\{\begin{array}{cc}\mathrm{<span\; class="notranslate">\; 1</span>}& \mathrm{<span\; class="notranslate">\; eval</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; query</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; term</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ></span>\mathrm{<span\; class="notranslate">\; eval</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; query</span>}<span\; class="notranslate">\; )</span>\\ \mathrm{<span\; class="notranslate">\; 0</span>}& \mathrm{<span\; class="notranslate">\; eval</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; query</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; term</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&le;</span>\mathrm{<span\; class="notranslate">\; eval</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; query</span>}<span\; class="notranslate">\; )</span>\end{array}\right\}$ Among them, eval() is the evaluation index function used to evaluate the retrieval performance, eval(query+term) is the score of the evaluation index function eval() when adding the candidate extended vocabulary term to the query query, and eval(query) is The score of the evaluation index function when evaluating the query query; the label marked as 1 indicates that the candidate extended vocabulary is related to the query query; the label marked as 0 indicates that the candidate expanded vocabulary is not related to the query query; The feature extraction of candidate extended vocabulary is to extract the distribution information of candidate extended vocabulary, the distribution information of candidate vocabulary in biomedical resources, and the candidate extended vocabulary from the biomedical resources and the first N query result documents returned by the query in the query pool. and the correlation information of the original query to prepare for the training ranking model, and after extracting multiple features of the same candidate extended vocabulary, normalize all feature values to control all feature values in [0,1] On the interval, the normalization process is as follows: $\mathrm{<span\; class="notranslate">\; newFeatureValue</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; oldFeatureValue</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; min</span>}\mathrm{<span\; class="notranslate">\; value</span>}}{\mathrm{<span\; class="notranslate">\; max</span>}\mathrm{<span\; class="notranslate">\; value</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; min</span>}\mathrm{<span\; class="notranslate">\; value</span>}}$ Among them, minValue and maxValue are the minimum and maximum values of a feature, respectively; S4. Training step of the candidate extended vocabulary sorting model: according to the relevance degree labeling and various features of the candidate expanded vocabulary, use the word grouping sorting algorithm to train to obtain the weight value of each feature. The specific steps are: select one that is marked as relevant in step S3 Candidate extended vocabulary and a number of candidate extended vocabulary marked as irrelevant form a word group, and select several such word groups as training samples; randomly assign initial weights to the features of each candidate extended word, and use the feature weighted score to evaluate each According to the ranking results of each word group, the overall ranking loss is calculated, and the weight of each dimension feature is dynamically adjusted according to the gradient value of the loss function, where the ranking loss is: Among them, NumSample is the number of candidate extended vocabulary groups in the word grouping, and loss _i is the loss value of each word grouping. The loss value is obtained by calculating the sorting position of the related expanded vocabulary. The higher the sorting position, the smaller the corresponding loss value; Iterate the previous process in a loop until the overall loss value is less than a certain threshold or reaches the specified number of iterations, and the training is completed, and the final selected feature value is used as the sorting model for the training completion; The online inquiry phase includes the following steps: S5. Online search engine query and extraction steps: For the new query submitted by the user online, retrieve the first N1 query results; extract the professional vocabulary and its various features in the first N1 search results according to biomedical resources, where N1 is a natural number; S6. Online Candidate Extended Vocabulary Extraction and Feature Extraction and Scoring Steps: According to the biomedical resources, use the candidate extended vocabulary extraction method in the offline stage S2-S3 and the feature extraction method for the candidate extended vocabulary to search for the first N1 search results In the online query stage, the professional vocabulary and its various features are extracted to obtain the candidate extended vocabulary in the online query stage, and the extracted features are used to measure the importance of the candidate extended vocabulary in the extended query; according to the feature weight obtained in step S4 training, the online query Stage candidate extended vocabulary is scored, and K1 candidate extended vocabulary with the highest score is selected to be added to the new query submitted online as an extended query, where K1 is a natural number; For a candidate extended vocabulary in the online query stage that is annotated and extracted using biomedical resources, its score is Where FeatureNum is the total number of features, ai is the weight value of the i-th feature in the ranking model, and feature _i (term) is the feature value of the i-th feature corresponding to the candidate extended vocabulary term in the online query phase; Rank them according to the scores of the candidate extended vocabulary in the online query stage, and select the top K1 candidate extended vocabulary in the online query stage as the extended vocabulary to add to the new query submitted online, the added candidate extended vocabulary in the online query stage is in The weights in the extended query can be expressed as $\mathrm{weight}={\mathrm{\&Sigma;}}_{i=1}^{\mathrm{count}}{\mathrm{weight}}_i\&CenterDot;\mathrm{feature}{e}_i+\mathrm{sign}\&CenterDot;\mathrm{weight}{t}_{\mathrm{original}},$ Wherein sign is a sign function, sign=1 when the candidate extended vocabulary in the online query stage appears in the new query submitted online, otherwise sign=0, weight _original is the weight value of the new query submitted online in the extended query; S7. Step of returning query results: performing retrieval according to the extended query, and returning the retrieval results to the user. 2. a kind of biomedical literature retrieval method based on word grouping sorting algorithm according to claim 1, it is characterized in that, in step S2, the weighted sum of the number of occurrences of professional vocabulary in described query result document is Among them, count _i is the number of times the word appears in the i-th document, and d(i) is the decay factor of the i-th document. 3. a kind of biomedical literature retrieval method based on word grouping sorting algorithm according to claim 1, is characterized in that, in step S3, evaluation index function eval () is average accuracy rate function, namely: ${\mathrm{<span\; class="notranslate">\; eval</span>}}_{\mathrm{<span\; class="notranslate">\; MAP</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{{\mathrm{<span\; class="notranslate">\; RelDoc</span>}}_{\mathrm{<span\; class="notranslate">\; query</span>}}}<span\; class="notranslate">\; \&Center\; Dot;</span>{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{{\mathrm{<span\; class="notranslate">\; RelDoc</span>}}_{\mathrm{<span\; class="notranslate">\; query</span>}}}\frac{\mathrm{<span\; class="notranslate">\; i</span>}}{\mathrm{<span\; class="notranslate">\; rank</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span>}$ Wherein, RelDoc _query is the number of related documents of a given query query, and rank(i) indicates the position of the i-th related document in the ranking list of document results. 4. a kind of biomedical literature retrieval method based on word grouping sorting algorithm according to claim 1, is characterized in that, in step S1, when there is no historical query record, by constructing biomedical query and retrieval method The method is to manually obtain the records of the query and its results; the retrieval method uses a vector space model, a BM25 retrieval model or a language model based on different smoothing methods. 5. a kind of biomedical literature retrieval method based on word grouping sorting algorithm according to claim 1, is characterized in that, loss value is in the step S4: Wherein, rank _i is the ranking position of the relevant candidate expansion words in the word grouping list. 6. A biomedical literature retrieval method based on word grouping and sorting algorithm according to claim 1, wherein the biomedical resources refer to dictionaries or knowledge bases containing biomedical professional vocabulary. 7. A kind of biomedical literature retrieval method based on word grouping sorting algorithm according to claim 1, is characterized in that, the feature of described candidate expanded vocabulary comprises the frequency TF that candidate expanded vocabulary appears in result document, candidate expanded vocabulary The TF-IDF value of , the number of documents in which the candidate extended vocabulary co-occurs with the original query, the number of times the candidate extended vocabulary and the original query co-occur in the same text window, the number of candidate extended vocabulary in biomedical resources, and the biomedical In the resources, the number of terms and concepts that contain the candidate extended vocabulary and the inclusion relationship between the terms and concepts of biomedical specialty. 8. A biomedical literature retrieval system based on word grouping sorting algorithm, is characterized in that, comprises off-line training part and on-line retrieval part; Described off-line training part comprises following part: Search engine query extraction module: used to extract multiple sets of queries and the first N query result documents obtained in each query according to the historical query records of the search engine; and collect the queries and query result documents into a query pool, where N is a natural number; Candidate Extended Vocabulary Extraction Module: It is used to extract professional vocabulary from the first N query result documents obtained by the search engine query extraction module by using the inherent resources in the biomedical field when a user query is given, and the professional vocabulary in the The frequency of occurrence in the query result document or the weighted sum of the number of occurrences is recorded; according to the number of occurrences of each professional word in the query result document or the weighted and descending order of the number of occurrences, select the M professional words with the highest frequency of occurrence as candidate extensions Vocabulary, where M is a natural number; Feature extraction and labeling module of candidate extended vocabulary: used to extract features related to the candidate extended vocabulary obtained in the candidate extended word extraction module, and mark the correlation of candidate extended vocabulary according to the impact of candidate extended vocabulary on retrieval performance degree; Candidate extended vocabulary sorting model training module: used to use the word grouping sorting algorithm to train the vocabulary sorting model to obtain the weight value of each feature of the candidate extended vocabulary after extracting the features of the candidate extended vocabulary and marking the degree of relevance of the candidate extended vocabulary; The online search section includes: Query reconstruction module: used for professional vocabulary extraction and candidate extended vocabulary scoring in new queries; including online search engine query extraction module, online candidate extended vocabulary extraction and feature extraction and scoring modules, in which the online search engine query extraction module is used Based on the new query submitted by the user online, the first N1 query results are retrieved; the professional vocabulary and its various features in the first N1 search results are extracted according to biomedical resources, where N1 is a natural number. The online candidate extended vocabulary extraction and feature extraction and scoring module uses the weight value of the candidate extended vocabulary output by the vocabulary ranking model to calculate the corresponding weight, and adds it to the original query to obtain the extended query; Query result return module: used to return the result document retrieved by the extended query to the user.