CN107247745B - A kind of information retrieval method and system based on pseudo-linear filter model - Google Patents

Abstract

The present invention provides an information retrieval method based on a pseudo-relevance feedback model, which includes merging the word relevance into the pseudo-relevance feedback model to realize information retrieval, including generating query expansion words in a pseudo-relevance document collection, respectively generating candidate extensions The query expansion words characterized by the importance of the word and the query expansion words characterized by the correlation between the candidate expansion words and the query subject words are combined into the original query expansion words to complete the final information retrieval; When querying extended words characterized by the correlation between the expanded word and the query subject word, the kernel function is used to calculate the correlation between the query word and the candidate word appearing in different positions in the document. The present invention can not only highlight the distribution of query words and candidate words, select candidate words with a higher degree of correlation with query subject words, but also locate more accurate candidate words, improve extended query and final retrieval due to the additional correlation information. accuracy.

Description

An information retrieval method and system based on a pseudo-correlation feedback model

technical field

The invention belongs to the technical field of information retrieval, and in particular relates to an information retrieval method and system which integrates the kernel function word correlation into a pseudo-correlation feedback model.

Background technique

In the age of information competition, browsing and obtaining desired information with the help of search engines is an important part of people's daily life. However, the extremely rich network resources and the rapid expansion of the total amount of information make it difficult for users to efficiently and accurately obtain and identify important information. Information processing technology urgently needs a more effective theory and method to deal with the growing mass of data. As a classic text processing technology, information retrieval can adapt to this requirement and quickly become a research hotspot in the current information processing research field.

Information Retrieval (Information Retrieval) refers to the process and technology of organizing information in a certain way and finding relevant information according to the needs of information users. The information retrieval process can be simply described as: the user organizes a query string and submits it to the information retrieval system according to his information needs, and the information retrieval system retrieves a subset of documents related to the query from the document set and returns them to the user. Specifically, given a set of specific query topics, through a certain information retrieval model, the relevance calculation is performed on all the documents in the target and the query topics, and each document is returned in order of scores from large to small, The higher the document in the returned results, the more relevant the document is to the query subject. After nearly half a century of research and development, some effective information retrieval models have been proposed and gradually applied to related systems. Among them, the retrieval models with relatively large influence include: Boolean logic model, vector space model, probability model, language model and the retrieval model based on supervised learning proposed recently.

In the actual information retrieval application, there is a certain deviation between the user's query request and the query result fed back by the system, which causes the performance of the retrieval system to decline. Therefore, information retrieval is often an iterative process, and users often need to go through multiple query adjustments to obtain satisfactory retrieval results. Query expansion technology, by expanding and reconstructing the user's initial query, better solves the problems of mismatch between user query words and document words and incomplete user expression, so it is widely used in the field of information retrieval. To put it simply, query expansion means that before searching, the retrieval system automatically expands the synonyms or near synonyms of the keywords in the user query according to the extended vocabulary to form a new query, and then performs the retrieval.

The appearance of pseudo-correlation feedback is to make the retrieval system more effective, so that the retrieval results can better meet the user's query request. The main mechanism is that the system defaults that the results retrieved by itself contain a large number of documents related to the subject of the user's query, and the first N articles are taken out as relevant documents to adjust or expand the query.

Usually, there are many factors that affect the performance of a retrieval system, the most critical of which is the information retrieval strategy, including the representation method of documents and query conditions, the matching strategy for evaluating the relevance of documents and queries, the sorting method of query results, and relevant feedback from users. mechanism etc.

With the rapid development of the Internet and the accumulation of massive information, the accuracy of information search has become the first point of concern for all users. Now it is becoming more and more difficult to find what users want through information retrieval tools. At the same time, various information Excessive flooding makes users have to spend more time identifying which information is valuable to users. The common problem of existing information retrieval methods is that the average retrieval accuracy is not high. Even the average accuracy of the current best retrieval model is only 30%, and there is still a long way to go to improve the information retrieval accuracy. However, information retrieval has penetrated into all aspects of human life. Most people use Baidu, Google and other search tools to find various information and solve various practical problems every day. In 2010, the number of web search requests in China reached more than 60 billion times, and in 2016, the number of search requests per day on Baidu alone reached 6 billion times. With such a large number of search requests, the average accuracy of information retrieval is A one percent improvement will save a lot of time and effort in getting the information you need, and the value it yields is extraordinary. Major Internet companies are also constantly pursuing lower-cost and more efficient information retrieval technologies.

Contents of the invention

The problem to be solved by the present invention is to optimize query expansion to finally improve the average precision of retrieval.

The present invention provides an information retrieval method based on a pseudo-relevance feedback model, which integrates word relevance into the pseudo-relevance feedback model to realize information retrieval, including generating query expansion words in a pseudo-relevance document collection, respectively generating candidate extensions The query expansion words characterized by the importance of the word and the query expansion words characterized by the correlation between the candidate expansion words and the query subject words are combined into the original query expansion words to complete the final information retrieval; When querying extended words characterized by the correlation between the expanded word and the query subject word, the kernel function is used to calculate the correlation between the query word and the candidate word appearing in different positions in the document.

Moreover, the word correlation is fused into the pseudo-correlation feedback model to realize information retrieval, and the implementation method is as follows,

When a user submits a query subject, the query subject is preprocessed to obtain the query keyword Q, D is all target documents, N _D is the total number of documents in the target document set D, and the query keyword Q and target are calculated through the preset retrieval weight model The score of each document in the document collection D is arranged according to the score results from high to low to obtain the first query result; assume that the first N documents in the target document collection D are taken out according to the pseudo-relevant feedback method as the pseudo-related document collection D _1. Perform the following steps when selecting query expansion words,

Step 1, take all the words in each document in the pseudo-related document collection D ₁ as extended candidate words, and calculate the importance score of each extended candidate word t _j in the document d _i of the pseudo-related document collection D ₁ Get the importance vector of each document d _i as follows,

Among them, i=1,2,3...,N, j=1,2,3...,n;

Calculate the importance score vector of the extended candidate word in all documents as follows,

Will After the importance _score of each extended candidate word in The corresponding extended candidate words are selected to form the importance query expansion word set Q ₁ , and the polynomial V ₁ is used to represent each word in the importance query expansion word set Q ₁ and the corresponding importance score of the word;

Step 2, take all the words in each document in the pseudo-related document set _D1 as extended candidate words, and use the kernel function to calculate each extended candidate word _tj and query keyword Q in the document d _i according to the co-occurrence position and frequency respectively Relevance Score for Get the correlation vector of each document d _i as follows,

Among them, i=1,2,3...,N, j=1,2,3...,n;

Calculate the vector of relevance scores of the extended candidate across all documents as follows,

Will After taking out the correlation score of each extended candidate word, it is sorted in descending order, and the top n ₁ values with the largest score are in the Corresponding expansion candidate words are selected to form the relevance query expansion word set Q ₁ ′, and the polynomial V ₁ ′ is used to represent each word in the query expansion word set Q _{1 ′} and the correlation score corresponding to the word;

Step 3, normalize the polynomials V ₁ and V ₁ ' obtained in steps 1 and 2 and then perform linear combination to obtain a new query word polynomial V as follows,

V＝(1-γ)×||V ₁ ||+γ×||V ₁ '||

Among them, ||X|| represents the normalization operation on the vector X, and γ is the adjustment factor;

Step 4: According to the query word polynomial V obtained in step 3, sort the coefficients of each term from large to small, and take out the first n ₁ terms with the largest coefficients to obtain a new set of expanded words

Step 5, set the query keyword Q to include the query word q _s , s=1,2,3...,m, express the query keyword Q as a polynomial V _Q , and set the coefficient value of each query word to 1.0; set step 4 The set of expanded words obtained from Expressed by the polynomial V',

The query polynomial V _Q and the query expansion word polynomial V' are normalized and linearly combined to obtain a new query word polynomial K as follows,

K＝α×||V _Q ||+β×||V'||

Among them, α and β are adjustment factors;

Step 6: Obtain a new query keyword set Q' according to the query word polynomial K obtained in step 5, use the weight corresponding to each query word in the query word polynomial K in the new query keyword set Q' and Q', and use the preset The retrieval weight model is used for the second information retrieval, and the obtained query results are taken as the final information retrieval results.

Also, in step 1, the importance score Use TFIDF, BM25 or RM3 to obtain.

Moreover, in step 2, calculate the correlation score between each extended candidate word t _j and the query keyword Q in the document d _i The implementation is as follows,

Let t _r and q _s co-occur in some document d _i , expressed as Calculated as follows,

in, Indicates the correlation between t _r and q _s in document d _i , Indicates that in document d _i the co-occurrence frequency of Indicates that in document d _i The co-occurrence anti-document frequency of ;

Calculate the correlation between t _r and the query keyword Q in the document d _i ,

Also, in document d _i The co-occurrence frequency of is calculated as follows,

Among them, M and L represent the number of occurrences of t _r and q _s in document d _i respectively, Indicates the k1th t _r appearing in document d _i , Indicates the k2th q _s appearing in document d _i , k1=1,2,3...,M, k2=1,2,3...,L; refers to the embodiment of the kernel function and proximity of the location.

Also, the kernel function is a Gaussian function or a trigonometric function.

Moreover, when the kernel function is a Gaussian function, the calculation is as follows,

Among them, p _t and p _q represent respectively and In the positional value in the document, σ is the tuning parameter.

Also, in document d _i The co-occurrence anti-document frequency of Calculated as follows,

in, means when hour, The total number of co-occurrences in document d _i .

Moreover, the preset retrieval weight model is based on a vector space model, a probability model or a language model.

The present invention also provides an information retrieval system based on a pseudo-correlation feedback model, which includes a computer or a server, and executes the above-mentioned method on the computer or server.

According to the information retrieval method provided by the present invention, which integrates the kernel function word correlation information into the pseudo-relevance feedback model, it can overcome the deficiency that the traditional pseudo-correlation feedback model only considers word frequency information. In addition, the correlation between query words and candidate words that appear in different positions in the document is calculated by the kernel function, which can not only highlight the distribution of query words and candidate words, but also select candidate words that are more relevant to the query subject words. Because of the additional relevance information, more accurate candidate words can be located, and the average precision of extended query and final retrieval can be improved. The experimental results of comparative experiments on multiple international information retrieval evaluation standard data sets with the best models in the world show that the information retrieval method that integrates the word relevance information of the kernel function into the pseudo-relevance feedback model provided by the present invention method, achieved a significant improvement in retrieval accuracy and reached the international leading level.

Description of drawings

FIG. 1 is a flowchart of a complete information retrieval process in an embodiment of the present invention.

Specific implementation method

The core problem to be solved by the present invention is to propose a kernel function to reflect the distribution of user query words and document candidate words and the degree of correlation between the two, and to use this degree of correlation as an additional weight in the pseudo In the relevant feedback model, query expansion is implemented to improve the accuracy of retrieval.

The following describes in detail the information retrieval method of the present invention, which integrates the word correlation degree of the kernel function into the pseudo-correlation feedback model, in conjunction with the accompanying drawings and embodiments.

Aiming at the unreasonable assumption of vocabulary independence in the classical method, the present invention proposes to take the correlation between words into consideration. Through the effective use of some statistical information of the data in the document collection (such as context information and other information reflecting the relationship between vocabulary collocations), and related technical solutions combined with query conditions to obtain vocabulary that can reflect the subject of the query conditions and be triggered by the query conditions, That is to use this information to more accurately capture the user's information needs.

The kernel function adopted in the method of the present invention is to project the linearly inseparable data in the original coordinate system to another space with Kernel, so as to make the data linearly separable in the new space as much as possible. In the method of the present invention, it will be used to evaluate the degree of correlation between two words in a document.

Referring to accompanying drawing 1, the procedure of the embodiment is, when the user searches according to the relevant query subject:

The information retrieval system will establish a query index based on the target document collection. When the user submits a relevant query topic, the system will preprocess the query topic into query keywords Q (Q is a collection, generally including multiple subject words q ₁ , q ₂ , q _3, etc.), D is all target documents, and N _D is the total number of documents in the target document collection D. Subsequently, the retrieval system will calculate the query keyword Q and the score of each document in the document collection D through a certain preset retrieval weight model (such as TFIDF, BM25, RM3, etc.), and arrange the score results from high to low to obtain The result of the first query. According to the principle of pseudo-relevance feedback, the retrieval system will take out the first N (in a large number of relevant research literature, N is generally 10, 20 or 30) documents in the first query result documents of the document collection D as pseudo-relevance For the document collection D ₁ , N is less than or equal to N _D , and those skilled in the art can preset values. When the retrieval system obtains the pseudo-related document set D ₁ generated by the first query, and selects query expansion words, the following steps are performed,

Step 1, respectively calculate the importance score of all the words (i.e. extended candidate words) in each document in the pseudo-related document collection _D1 , the importance score can be obtained by the word frequency of the word and the reverse document word frequency (such as TFIDF, BM25, RM3, etc.), and then accumulate the same word importance scores in different documents in the form of word vectors and divide by the number of documents N in _D1 to obtain the importance score vectors of all extended candidate words Arrange the scores of the elements in the vector in descending order, and take out the first n ₁ (n ₁ is generally a value of 10, 20, 30 or 50, and those skilled in the art can preset the value) scores in The corresponding word in , get the importance expansion candidate word set Q ₁ , and a polynomial V ₁ can be used to represent each word in the set Q ₁ and the corresponding importance score of the word.

In the present invention, each document in the set of N pseudo-related documents _D1 is regarded as a bag-of-words model, which is expressed in the form of a word vector, and the correlation vector formula of the i-th document is as follows.

In the above formula, Represents the word vector expression of the i-th document (i=1,2,3...,N)d _i in the pseudo-related document collection _D 1 , t ₁ , t ₂ , t ₃ ,..., t _n are pseudo-related document collections All words in all documents in D ₁ , n represents the total number of these words, that is, the number of all words in the pseudo-related document collection D ₁ ; Indicates the corresponding weight scores of t ₁ , t ₂ , t _{3 , .} The importance score of a word is obtained by calculating the word frequency of the word and the inverse document word frequency (such as TFIDF, BM25, RM3, etc.). For example, when using TFIDF to calculate the importance of term t _j in document d _i ,

in, Refers to the importance score of a term t _j in document d _i (j=1,2,3...,n), TF(t _j ,d) refers to the frequency of term t _j appearing in document d _i ( times), N _D is the total number of documents in the target document set D, and df(t _j ) is the number of documents in the pseudo-related set D ₁ that contain the term t _j .

According to formula (2), each document d _i in the N documents can be expressed as a vector form of the importance of the corresponding word And accumulate and sum each document vector and then divide it by the total number of pseudo-related documents N to get the importance score vector of all terms in all documents As shown in formula (3):

Will After the importance score of each word in is taken out, it is sorted in descending order, and the top n ₁ values with the largest score are in The corresponding words are selected to form an importance query expansion word set Q ₁ . In order to facilitate subsequent calculations, a polynomial V ₁ is used to represent each word in the set Q ₁ and the corresponding importance score of the word, as shown in formula (4).

In formula (4), qh ₁ , qh ₂ , qh ₃ , ..., Indicates each specific expansion candidate word in Q ₁ (a total of n ₁ ), wh ₁ , wh ₂ , wh ₃ ,..., Indicates that the corresponding extended candidate words are in score in .

Step 2, by sequentially calculating the correlation score between all the words (i.e. extended candidate words) and the query word in each document in the pseudo-related document set _D1 , the correlation score is based on the query word and the extended candidate word in each document The position of is calculated by the kernel function, and then the scores of the same word in different documents are accumulated to obtain the correlation score vector of all extended candidate words and query words Arrange the scores of the elements in the vector in descending order, and take out the first n ₁ (n ₁ is generally 10, 20, 30 or 50) scores in Corresponding words in , get the set Q ₁ ' of candidate words for correlation expansion, here we use a polynomial V ₁ ' to represent the correlation score between each word in the set Q ₁ ' and the word.

For ease of explanation, given extended candidate word t _r and query word q _s , (where r=1, 2, 3..., n, n is the number of all words in the pseudo-related document collection D ₁ , s=1, 2,3...,m, m is the number of words in the set of query keywords Q), if t _r and q _s co-occur in a certain document d _i , expressed as Then they have a co-occurrence weight (ie correlation). Since t _r and q _s may appear in multiple places in a document, they cannot simply be The number of co-occurrences of represents the correlation between t _r and q _s in document d _i . In order to measure it more reasonably, the present invention further proposes the following formula:

In formula (5), Indicates the correlation between t _r and q _s in document d _i .

In formula (5), Indicates that in document d _i The co-occurrence frequency of , the specific calculation formula is as follows:

In formula (6), M and L represent the number of occurrences of t _r and q _s in document d _i respectively, Indicates the k1th t _r appearing in document d _i , Indicates the k2th q _s appearing in document d _i , k1=1, 2, 3..., M, k2=1, 2, 3..., L. Among them, Kernel() represents the kernel function. The kernel function is a kind of adjacent relationship between two words that can be measured by the position information of the word. When the co-occurrence position of the two words is closer, the adjacent relationship is stronger, that is, the correlation to a higher degree. Such as Gaussian functions, trigonometric functions, etc. are very effective in many scenarios. Example of It refers to the Gaussian kernel function (other kernel functions can also be used in specific implementation) to reflect and The proximity of the location, such as the formula (7):

Among them, p _t and p _q represent respectively and The position value in the document (that is, the sequence number of the word in the document, which is a positive integer), σ is an adjustment parameter, which is used to adjust the distribution of the Gaussian function, and the value range of σ is between 10 and 100. In the specific Preferably take 50 in the embodiment.

In formula (5), Indicates that in document d _i The co-occurrence anti-document frequency of is calculated as follows:

in, means when hour, The total number of co-occurrences in document d _i .

Equation (5) gives the correlation between t _r and q _s in document d _i Since q _s is a query term in the query keyword set Q, by formula (5), the correlation between t _r and the query keyword Q in the document d _i can be obtained, and the present invention uses To express, the specific calculation formula is as follows:

According to the formula (9), the i-th document d _i in the set of N pseudo-related documents D ₁ can be expressed in the form of the correlation vector between the corresponding extended candidate words and query words, that is, The specific formula is as follows.

Next, for each document relevance vector Carry out cumulative summation and then divide by the total number of pseudo-related documents N, and finally get the relevance score vector of all terms in all documents As shown in formula (11):

Will After taking out the correlation score of each word, sort it in descending order, and put the top n ₁ values with the largest score in Corresponding words are selected to form the set Q ₁ ' of relevance query expansion words. To facilitate subsequent calculations, the polynomial V ₁ ' is used to represent each word in the set Q ₁ ' and the correlation score corresponding to the word, as shown in formula (12).

In formula (12), qh ₁ ', qh' ₂ , qh ₃ ', ..., Indicate each specific expansion word in Q ₁ ' (n ₁ in total), wh′ ₁ , wh′ ₂ , wh′ ₃ , ..., Indicates that the corresponding expansion word is in score in .

Step 3: Normalize the query expansion term polynomials V ₁ and V ₁ ′ obtained in Step 1 and Step 2, and then linearly combine them to obtain a new query term polynomial V. The specific combination method is shown in formula (13).

V＝(1-γ)×||V ₁ ||+γ×||V ₁ '|| Formula (13)

In formula (13), ||X|| represents the normalization operation on the vector X. The purpose of normalization is to unify the dimension, that is, to standardize the value of each element in the vector to the interval [0,1.0]. It is convenient for subsequent parameter adjustment. Normalization can be realized in many ways. In this embodiment, the method of dividing by the maximum value is adopted, that is, the normalized value of each element is divided by the original value of the element by the maximum value of the element in the vector. For example, there is a vector [1,2,3,4] with 4 elements, and the maximum value of the element is 4, then the normalized result of dividing this vector by the maximum value method is That is [0.25,0.5,0.75,1], you can see that all the values in the original vector are normalized to the interval [0,1.0].

The adjustment factor γ in the formula (13) ranges from 0 to 1.0, and its function is to balance the importance score of the extended word and the correlation score between the extended word and the query word. In specific applications, it can be adjusted in advance Use the test data to test on the target document set that needs to be applied to get the optimal value of γ.

Step 4, according to the polynomial V in step 3, sort the coefficient (comprehensive weight score) of each term from large to small, and take out the first n+ ₁ terms with the largest coefficient to obtain a new extended word set That is, the final set of query expansion words.

Step 5, express the original query keyword set Q as a polynomial V _Q , each item in the polynomial V _Q is each query word q _s in Q, s=1,2,3...,m, each item's If the coefficient value is set to 1.0, it can be expressed as

V _Q ＝1.0×q ₁ +1.0×q ₂ +1.0×q ₃ +...+1.0×q _m formula (14)

Next, the expanded word set obtained in step 4 Also represented by a polynomial V', each term in the polynomial V' is For each specific query expansion term in , the coefficient of each item (term) is the corresponding value of the term in the polynomial V in step 4,

Among them, q' ₂ , q' ₃ ,..., express Each specific expansion word in (a total of n ₁ ), w' ₁ , w' ₂ , w' ₃ ,..., Indicates the score of the corresponding expanded word in the query word polynomial V.

The query polynomial V _Q and the query expansion word polynomial V' are normalized and linearly combined again to obtain a new query word polynomial K. The specific combination method is shown in formula (16).

K＝α×||V _Q ||+β×||V'|| formula (16)

In the formula (16), the normalization method consistent with step 3 is adopted. The adjustment factor α in the formula generally takes a fixed value of 1.0, and the value range of the adjustment factor β is 0 to 1.0. Its function is to balance the original The weight between the query word and the extended query word can be set as an empirical value during specific implementation.

Step 6, according to step 5, a new query keyword set Q' can be obtained, and each query word in Q' is each term in the query word polynomial K. Use the weight of each query term in the query term polynomial K in the new query keyword set Q' and Q' to perform the second information retrieval (the same retrieval model as the first retrieval), that is, calculate Q' again and the score of each document in the target document set D, the query result obtained is the final information retrieval result.

When performing the second search, the query word is the newly generated query keyword set Q', and when calculating the score of the query word and each document, the weight of each query word is the coefficient of the word in the query word polynomial K , while each query term is given a weight of 1.0 during the first retrieval.

During specific implementation, those skilled in the art can use software technology to realize the automatic operation of the above process. Correspondingly, if an information retrieval system based on a pseudo-relevance feedback model is provided, including a computer or a server, and the above process is executed on the computer or server to integrate the word correlation into the pseudo-relevance feedback model to realize information retrieval, it should also be included in the present invention within the scope of protection.

For example, the development environment for information retrieval is Java and Python development environments, and the development support library is Lucene.

The information retrieval framework can be based on vector space models, probability models, language models and other pseudo-relevant feedback information retrieval frameworks.

In order to verify the actual effect of the method of the present invention, a comparative experiment has been done on a plurality of standard data sets, and the comparative experiment is divided into two groups, one group adopts the standard Rocchio pseudo-correlation feedback information retrieval model, and the other group adopts the combination of the method of the present invention The Rocchio pseudo-relevance feedback information retrieval model is abbreviated as KRC. This experiment uses six standard international datasets, including AP88-89, AP90, DISK1&2, DISK4&5, WT2G, and WT10G. The information of these datasets is shown in the following table (Table 1):

dataset name total number of documents size Query subject number Number of query topics AP90 78,321 0.23Gb 51-100 50 AP88-89 164,597 0.50Gb 51-100 50 DISK1&2 741,856 2.03Gb 51-200 150 DISK4&5 528,155 1.85Gb 301-450 150 WT2G 247,491 2.14Gb 401-450 50 WT10G 1,692,096 10Gb 451-550 100

Table 1 Basic information of the six datasets

In the comparison experiment, the kernel function in the method of the present invention is adopted to select a Gaussian kernel function (other kernel functions can also be used), and the value of σ in the Gaussian kernel function is 50. In order to make the experiment more fair, the number of query expansion words N1 was selected in four cases: 10, 20, 30 and 50 respectively. The experimental results in different cases are shown in the following table (Table 2):

Table 2 Comparison of average precision (MAP) between Rocchio and KRC models on six standard datasets

In table 2, the Rocchio model in the second column does not adopt the method of the present invention, and the KRC model is the Rocchio model that has adopted the method of the present invention, and MAP is the average precision of retrieval result, can observe from the table, the method of the present invention The retrieval accuracy improvement effect on the Rocchio pseudo-correlation model is remarkable, indicating that the technical solution of the present invention is effective.

Claims (9)

1. An information retrieval method based on a pseudo-relevance feedback model, characterized in that: the word relevance is fused into the pseudo-relevance feedback model to realize information retrieval, including generating query expansion words in a pseudo-relevance document collection, respectively generating the following The query expansion words characterized by the importance of the candidate expansion words and the query expansion words characterized by the correlation between the candidate expansion words and the query subject words are combined into the original query expansion words to complete the final information retrieval; generate When the query expansion word is characterized by the correlation between the candidate expansion word and the query subject word, the kernel function is used to calculate the correlation between the query word and the candidate word appearing in different positions in the document; Said fusion of word relevance into the pseudo-correlation feedback model to realize information retrieval, the implementation is as follows, When a user submits a query subject, the query subject is preprocessed to obtain the query keyword Q, D is all target documents, N _D is the total number of documents in the target document set D, and the query keyword Q and target are calculated through the preset retrieval weight model The score of each document in the document collection D is arranged according to the score results from high to low to obtain the first query result; assume that the first N documents in the target document collection D are taken out according to the pseudo-relevant feedback method as the pseudo-related document collection D _1. Perform the following steps when selecting query expansion words, Step 1, take all the words in each document in the pseudo-related document collection D ₁ as extended candidate words, and calculate the importance score of each extended candidate word t _j in the document d _i of the pseudo-related document collection D ₁ Get the importance vector of each document d _i as follows, Among them, i=1,2,3...,N, j=1,2,3...,n; Calculate the importance score vector of the extended candidate word in all documents as follows, Will After the importance _score of each extended candidate word in The corresponding extended candidate words are selected to form the importance query expansion word set Q ₁ , and the polynomial V ₁ is used to represent each word in the importance query expansion word set Q ₁ and the corresponding importance score of the word; Step 2, take all the words in each document in the pseudo-related document set D ₁ as extended candidate words, and use the kernel function to calculate each extended candidate word t _r and the query keyword Q in the document d _i according to the co-occurrence position and frequency respectively Relevance Score for Get the correlation vector of each document d _i as follows, Among them, i=1,2,3...,N, r=1,2,3...,n; Calculate the vector of relevance scores of the extended candidate across all documents as follows, Will After taking out the correlation score of each extended candidate word, it is sorted in descending order, and the top n ₁ values with the largest score are in the Corresponding extended candidate words are selected to form a relevance query expansion word set Q′ ₁ , and the polynomial V ₁ ′ is used to represent each word in the query expansion word set Q′ ₁ and the corresponding correlation score of the word; Step 3, normalize the polynomials V ₁ and V ₁ ' obtained in steps 1 and 2 and then perform linear combination to obtain a new query word polynomial V as follows, V＝(1-γ)×||V ₁ ||+γ×||V ₁ '|| Among them, ||X|| represents the normalization operation on the vector X, and γ is the adjustment factor; Step 4: According to the query word polynomial V obtained in step 3, sort the coefficients of each term from large to small, and take out the first n ₁ terms with the largest coefficients to obtain a new set of expanded words Step 5, set the query keyword Q to include the query word q _s , s=1,2,3...,m, express the query keyword Q as a polynomial V _Q , and set the coefficient value of each query word to 1.0; set step 4 The set of expanded words obtained from Expressed by the polynomial V', The query polynomial V _Q and the query expansion word polynomial V' are normalized and linearly combined, and the new query word polynomial K is as follows, K＝α×||V _Q ||+β×||V'|| Among them, α and β are adjustment factors; Step 6: Obtain a new query keyword set Q' according to the query word polynomial K obtained in step 5, use the weight corresponding to each query word in the query word polynomial K in the new query keyword set Q' and Q', and use the preset The retrieval weight model is used for the second information retrieval, and the obtained query results are taken as the final information retrieval results. 2. The information retrieval method based on the pseudo correlation feedback model according to claim 1, characterized in that: in step 1, the importance score Use TFIDF, BM25 or RM3 to obtain. 3. according to the described information retrieval method based on pseudo correlation feedback model of claim 1, it is characterized in that: in step 2, calculate the relevancy score of each expanded candidate word t _r and query keyword Q in document d _i The implementation is as follows, Let t _r and q _s co-occur in some document d _i , expressed as Calculated as follows, in, Indicates the correlation between t _r and q _s in document d _i , Indicates that in document d _i the co-occurrence frequency of Indicates that in document d _i The co-occurrence anti-document frequency of ; Calculate the correlation between t _r and the query keyword Q in the document d _i , 4. according to the described information retrieval method based on pseudo correlation feedback model of claim 3, it is characterized in that: in document d _i The co-occurrence frequency of is calculated as follows, Among them, M and L represent the number of occurrences of t _r and q _s in document d _i respectively, Indicates the k1th t _r appearing in document d _i , Indicates the k2th q _s appearing in document d _i , k1=1,2,3...,M, k2=1,2,3...,L; refers to the embodiment of the kernel function and proximity of the location. 5. The information retrieval method based on the pseudo correlation feedback model according to claim 4, characterized in that: the kernel function is a Gaussian function or a trigonometric function. 6. according to the described information retrieval method based on pseudo correlation feedback model of claim 5, it is characterized in that: when kernel function is Gaussian function, calculate as follows, Among them, p _t and p _q represent respectively and In the positional value in the document, σ is the tuning parameter. 7. The information retrieval method based on the pseudo-correlation feedback model according to claim 4, characterized in that: in the document d _i The co-occurrence anti-document frequency of Calculated as follows, in, means when hour, The total number of co-occurrences in document d _i . 8. According to claim 1 or 2 or 3 or 4 or 5 or 6 or 7 described information retrieval method based on pseudo correlation feedback model, it is characterized in that: described preset retrieval weight model is based on vector space model, probability model or language models. 9. An information retrieval system based on a pseudo-relevant feedback model, characterized in that it comprises a computer or a server, and the method according to claims 1 to 8 is executed on the computer or server.