Multi-view fusion-based result diversification search method and system

Abstract

The application provides a multi-view fusion-based result diversification search method and system. The method first calculates the initial representation of a webpage document and a subtopic; then learns the attention relationship between the document and the subtopic through an attention mechanism under a global perspective, calculates the coverage degree representation, novelty representation and coverage degree representation of the document to the subtopic; then, from a local perspective, considers the dependency relationship between candidate documents, models the diversity features of the document set under the local perspective according to the local changes of the selected document list, and learns the local time sequence correlation representation of the document; then, the relevance between the document and the search statement is modeled, and the relevance representation between the document and the search statement is calculated; finally, a search result diversification ranking model is constructed, and the search result is returned. The application can effectively meet the diversification search needs of users, and can to some extent solve the problems of high content redundancy and single topic content bias of the returned results of the current traditional search.

Description

TECHNICAL FIELD

[0001] The present application relates to a result diversification search method, in particular to a result diversification search method based on multi-view fusion. BACKGROUND

[0002] The rapid development of the Internet has promoted the rapid growth of information, and the manual data retrieval method can not meet the daily needs of people. In order to alleviate information overload and speed up information acquisition, information retrieval technology emerges as the times require and develops rapidly. Search is one of the most important research directions in information retrieval, and the purpose of search is to retrieve web documents that meet user needs from a resource set according to the search statement initiated by the user. The result diversification search technology can better meet the search needs of users by mining the content diversity relationship of the document list, so that the homepage of the search result presents the characteristics of wider content theme coverage and more diversified views.

[0003] Unlike traditional search technology, result diversification search technology not only considers the relevance features between documents and search statements, but also considers the diversity relationship between documents. Unlike relevance modeling, diversity feature modeling considers that documents are not independent but mutually influenced, and considers not only the content theme relationship between documents and search subtopics, but also the content diversity relationship between documents, so the modeling degree is more difficult. In the field of diversification search, the heuristic method depends on the manually selected features, and the diversity feature representation is insufficient, and the parameters cannot be automatically learned, so the effect is not good. The supervised learning method depends on the manually labeled training data, can automatically optimize the parameters, and can more deeply and comprehensively mine the diversity relationship of the documents. Among them, the greedy modeling method under the local view ignores the content dependency between the candidate documents, does not consider the global association relationship of the documents, and can only get a local optimal solution, and cannot get a global optimal document result list. The modeling method under the global view considers the content association relationship between all documents, but loses the time sequence features of the document selection under the local view, that is, the influence of the selection of a document on the marginal information gain of the subsequent documents, so the search performance is limited. SUMMARY

[0004] The present application provides a result diversification search method and system based on multi-view feature fusion to solve the problems in the prior art.

[0005] Technical scheme: In order to achieve the above application purpose, the present application adopts the following technical scheme:

[0006] The result diversification search method based on multi-view fusion comprises the following steps:

[0007] (1) Information representation

[0008] By analyzing and understanding the context of the information, the initial representations of web page documents, search statements and their subtopics are calculated;

[0009] Step (1) includes the following specific steps:

[0010] (1-1) Input the set of web page documents P = [p1, p2, p3, ..., p N ], where p i Let N represent the i-th document and N represent the total number of documents.

[0011] (1-2) The PV-DBOW training mechanism of the Doc2Vec model is used to obtain document vectors containing word semantics and word order information. In the PV-DBOW training method, the i-th document represents p. i =[w i,1 ,w i,2 ,w i,3 ,...,w i,m ], where w i,j Let represent the j-th word of the i-th document, and m represent the total number of words in the document. Finally, given the text window size, a word is randomly sampled from the text window. By predicting the probability of this word, the document matrix representation is iteratively trained to obtain the document information representation. The goal of the training process is to maximize the probability of the predicted word. The calculation process is as follows:

[0012]

[0013] Where f(p) represents the calculation of the prediction probability of randomly sampled words in document p, and ω (v,z) Let P(ω) represent the z-th randomly sampled predicted word in the v-th text window, W be the set of text windows, v be a randomly sampled text window in W, V be the set of words in window v, and z be a randomly sampled word in V. (v,z) |d) represents generating predicted words ω for a given document p. (v,z) The normalized probability; its calculation process is as follows:

[0014]

[0015] Among them, y k Let ∑ represent the probability of predicting the k-th word before normalization. k exp(y k This represents summing the probabilities of all predicted words before normalization, and finally updating the document vectors using gradient descent to obtain the final document set representation. The formula for calculating the information representation of the document set is as follows:

[0016] D = Doc2Vec(P)

[0017] Wherein, D represents the final document set representation, Doc2Vec(·) represents the Doc2Vec model, P represents the web document set;

[0018] (1-3) Through the representation learning of information in the data set, the web document set P = [p1, p2, p3, …, p N ] is continuously iteratively updated and learned to obtain the final document set information representation D = [d1, d2, d3, …, d N ], when searching for information representation of search statements and sub-topics, first use the relevance search method to recall documents, then splice the top documents into a new document, and use the information representation result of the new document as the information representation vector of the corresponding search statement or sub-topic. In order to be in the same semantic space as the document representation, the search statement and sub-topic are also represented using the Doc2Vec model to obtain the information representation q of the search statement and the information representation S = [s1, s2, s3, … s k ];

[0019] (2) Global diversity representation calculation

[0020] Combined with the initial information representation of the web document, search statement and sub-topic obtained in step (1), the self-attention mechanism is used to calculate the novelty representation of the document and the sub-topic from the text perspective and the sub-topic perspective respectively, and the cross-attention mechanism is used to calculate the cross-attention representation of the fusion of the document and the sub-topic. Then, using the attention mechanism, the cross-attention representation of the web document is used as the query vector, and the cross-attention representation of the sub-topic is used as the key vector and the value vector to calculate the coverage representation of the web document to the sub-topic. Finally, the novelty representation of the web document and the coverage representation of the web document to the sub-topic are combined as the global diversity representation;

[0021] Step (2) includes the following specific steps:

[0022] (2-1) Novelty encoding stage, the last layer representation D l-1 = [d1, d2, d3, …, d N ] of the document set is used as the initial input of the self-attention mechanism, and the multi-head self-attention mechanism is used to update the representation between the documents to obtain the document representation D l of the current layer. After L-layer multi-head self-attention calculation, the document self-attention encoding matrix is obtained. Wherein, di represents the vector representation of the i-th document after the novelty encoder, and the multi-layer multi-head self-attention encoder divides the novelty representation into three parts,

[0023] First, the attention coefficient a l,t is calculated, and the calculation process is as follows:

[0024] al,t =Attn(D l-1 D l-1 D l-1 )

[0025] Among them, a l,t This represents the attention coefficient of the t-th group in the l-th layer, where query Q, key K, and value V are all document matrix representations D output by the multi-head self-attention mechanism of the previous layer. l-1 D l-1 The output of the (l-1)th hidden layer is concatenated with h sets of self-attention mechanism coefficients to obtain the multi-head self-attention coefficients (MHSA). l =[a l,1 ;a l,2 ;a l,3 ;…;a l,h ];

[0026] Secondly, the multi-head self-attention representation D of the l-th layer is calculated through a feedforward neural network, residual connections, and layer normalization operations. l =MultiSelfAttn l (S l-1 The calculation process is as follows:

[0027] X = LayerNorm(D) l-1 +MHSA l )

[0028] D l =LayerNorm(X+FeedForward(X))

[0029] Where X represents the input, MHSA l represents the multi-head self-attention coefficient, FeedForward(X) represents the feedforward neural network, and LayerNorm(·) represents the layer normalization operation;

[0030] Finally, a document matrix representation containing novelty information is learned through a multi-layer multi-head self-attention encoder. in This represents the final vector representation of the i-th document after it has been processed by the novelty encoder; its calculation process is as follows:

[0031] D enc =MultiSelfAttn L (MultiSelfAttn L-1 (…MultiSelfAttn1(D)))#

[0032] Among them, MultSelfAttn L This represents multi-head attention at layer L;

[0033] The calculation method of sub-topic novelty representation is consistent with the above method, the only difference is that the initial input is changed to sub-topic set information representation S=[s1, s2, s3, …, s k ], and finally the multi-layer multi-head self-attention encoded sub-topic novelty representation is obtained

[0034]

[0035] (2-2) Sub-topic coverage decoding stage, the document novelty representation and the sub-topic novelty representation obtained in step (2-1) are taken as input, the cross-attention is calculated between the sub-topic novelty representation and the document novelty representation to calculate the association degree between the document and the sub-topic, and the document cross-attention feature representation is obtained and the sub-topic cross-attention feature representation The calculation process is as follows:

[0036] D co =CrossAttention(D enc ,S enc )

[0037] S co =CrossAttention(S enc ,D enc )

[0038] Where CrossAttention(·) represents cross-attention, D enc represents the document novelty representation, and S enc represents the sub-topic novelty representation.

[0039] Taking the first layer calculation as an example, the specific calculation of cross-attention is as follows:

[0040] a t =Attention(D co ,S co ,S co )t∈[1,h]

[0041]

[0042] Where a t represents the attention coefficient of the tth attention head, and the multi-head attention coefficient MHA l is obtained by concatenating h attention coefficients hSimilar to the encoder in (2-1), the decoder obtains the multi-head attention representation of the document by calculating multi-head attention coefficients, layer normalization operations, residual connections, and a feedforward network, with ReLU as the activation function; subsequently, the decoder obtains the multi-angle and multi-level attention representation of the document pairs to sub-topics through an L-layer multi-head attention mechanism, the formal representation of which is as follows:

[0043] D dec =MultiAttn L (MultiAttn L-1 (…MultiAttn1(D co )))

[0044] D dec The final representation output by the decoder represents the degree to which the document covers the subtopics;

[0045] (2-3) By splicing the document novelty representation obtained in step (2-1) and the document coverage representation of subtopics obtained in step (2-2), the global diversity representation of the document is obtained.

[0046] (3) Calculation of local diversity characterization

[0047] Combining the initial representation of the web page document obtained in step (1), according to the document selection order in the selected web page document, the temporal information representation of the document is calculated using a long short-term memory network. The document graph is constructed, and the temporal information representation is used as the node representation of the document graph. The temporal association representation between documents is calculated through the graph attention mechanism as the local diversity representation.

[0048] Step (3) includes the following specific steps:

[0049] (3-1) The temporal relation representation containing the content of the preceding documents is obtained by learning through the LSTM model. The hidden state representation of the i-th document can be computed in the hidden layer of the LSTM. The calculation process is as follows:

[0050]

[0051] Where, d i This represents the initial input representation of the i-th document in the LSTM. This represents the temporal feature representation output after the i-th document is computed by LSTM;

[0052] (3-2) Construct a document graph by treating each document in the set as a node. The edge relationships of the graph are defined by the adjacency matrix E, which is an N×N matrix. ab This indicates the connection relationship between document node a and document node b. If e ab =1 indicates d a With db Connection, e ab =0 means d a With d b Not connected; its definition is as follows:

[0053]

[0054]

[0055] Considering that all documents in the set are related to the current search, and each document learns potential diversity features by relying on its relationship with all other documents, the document graph is set as a fully connected graph. Each time a document is selected, the document graph will be reconstructed, and the latest temporal features calculated in step (3-1) will be used as node representations.

[0056] (3-3) Use a graph attention network to update document nodes, document d i By aggregating the features of the previous layer and its own

[0057] The feature information of neighboring nodes is used to obtain the feature representation of this layer. The calculation process is as follows:

[0058]

[0059] in, N(d) represents the updated representation of the i-th document at level l. i ) represents the document node d i The set of all neighboring nodes, d g ∈N(d i ), σ(·) is a non-linear activation function, Aggregation(·) is an aggregation function, the number of layers in the graph attention network is set to l, and the updated representation of the i-th document at layer l is... That is, its local diversity is represented;

[0060] (4) Correlation characterization calculation

[0061] Combining the initial representations of web page documents, search statements and subtopics in step (1), the initial relevance representations between documents and subtopics are calculated using the document-dependent and search statement-dependent methods and the document-only method, respectively. The initial relevance representations are then processed by a multilayer perceptron to obtain the updated relevance representations. Considering the topic overlap and semantic redundancy between subtopics, the score obtained by multilayer perceptron for the subtopic novelty representation calculated in step (2) is multiplied by the updated relevance representation as the relevance coefficient to obtain the final relevance representation.

[0062] (5) Search result sorting module

[0063] The global diversity representation obtained in step (2), the local diversity representation obtained in step (3), and the relevance representation obtained in step (4) are combined to construct a search result diversification module, the global diversity representation, the local diversity representation, and the relevance representation are spliced and a document score is calculated through a multi-layer perception, and finally the top k documents with the highest scores are selected as the most diverse search result list and returned to the user;

[0064] Further, step (4) comprises the following specific steps:

[0065] (4-1) using the calculation methods dependent on the document and the search statement: TF-IDF, BM25, and LMIR, and the calculation method dependent only on the document content: PageRank, inlinks, and outlinks to calculate the initial relevance representation of the document and the subtopic Taking TF-IDF as an example, the initial relevance representation calculation process is as follows:

[0066]

[0067]

[0068] score tf-idf =[score tf,body ,score tf,anchor ,score tf,title ,score tf,URL ,score tf,doc ]

[0069] Wherein, represents the relevance representation of the i-th document and the subtopic set under the current search statement, represents the relevance representation of the i-th document corresponding to the k-th subtopic, score tf-idf represents the representation obtained by the TF-IDF method;

[0070] (4-2) the initial relevance feature D S rel After processing by the multi-layer perception, the document obtains a relevance score for each subtopic, and a total of k relevance scores are obtained as the updated relevance feature D score , and the calculation process is as follows:

[0071] D score =MLP(D S rel )

[0072] d i score =[di,1 score ,d i,2 score ,d i,3 score ,...,d i,k score ]

[0073] where MLP(·) denotes a multi-layer perceptron, d i,k score denotes the relevance score of the k-th subtopic in the i-th document;

[0074] (4-3) Concatenate the subtopic novelty representation S enc After the multi-layer perceptron processing, each subtopic corresponding to the search query gets a novelty score score i As the weight coefficient of the relevance feature, the weight coefficients of the k subtopics are represented as S w = [score1, score2, score3,..., score k ], multiply the weight coefficient matrix with the relevance feature, and get the final relevance feature representation D rel ; The calculation process is as follows:

[0075]

[0076]

[0077] where, denotes the relevance feature of the i-th document and the k-th subtopic, denotes the relevance feature of the i-th document, D rel denotes the relevance feature matrix representation of all candidate documents.

[0078] Further, step (5) includes the following specific steps:

[0079] (5-1) Concatenate the global diversity representation obtained in step (2), the local diversity representation obtained in step (3), and the relevance representation obtained in step (4), and the final feature representation of the i-th document is as follows:

[0080]

[0081] where, is the final feature representation of the i-th document, d i enc is the novelty representation of the i-th document, d i dec is the subtopic coverage representation of the i-th document, d iGAT is the local diversity representation of the i-th document, d i rel is the relevance feature representation of the i-th document;

[0082] (5-2) Using a multi-layer perception, the score of the document is calculated, and the document is ranked according to the score, and the calculation process is as follows:

[0083] Score = softmax(W o (σ(W F (D feature ))

[0084]

[0085] where W o ,W F is the parameter matrix of the MLP, D feature is the final matrix representation of the set of all candidate documents, and Score is the final score of the document. The documents are arranged in descending order of the score to obtain the optimal result diversified search ranking list.

[0086] Advantages: Compared with the prior art, the technical scheme adopted by the present application has the following advantages:

[0087] (1) The present application fuses diversified representation under global and local perspectives, which can overcome the shortcomings of traditional global diversity modeling that does not consider the influence of document selection on the marginal information of candidate documents, and solve the problem of not being able to obtain a globally optimal document result list caused by not adding global diversity relationship in local diversity modeling;

[0088] (2) The present application uses a local perception-based document time sequence correlation feature modeling method based on a document relationship graph, which can effectively utilize the document selection time sequence features, learn the diversity correlation between the document nodes and their neighbor document nodes through the relationship between the document nodes and their neighbor document nodes, and more deeply mine the local diversity features of the documents. BRIEF DESCRIPTION OF DRAWINGS

[0089] Figure 1 is the overall framework diagram of the algorithm of the present application;

[0090] Figure 2 is a global diversity modeling schematic diagram;

[0091] Figure 3 is a local diversity modeling schematic diagram;

[0092] Figure 4 is a system framework diagram. DETAILED DESCRIPTION

[0093] The technical scheme of the present application will be described in detail below.

[0094] The following is only one embodiment of the present invention. The present invention has many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art can make various corresponding changes and modifications according to the present invention. All such corresponding changes and modifications should fall within the protection scope of the appended claims.

[0095] This invention provides a result diversification search method based on multi-view fusion, the method comprising the following steps:

[0096] 1. Information representation

[0097] By analyzing and understanding the context of the information, the initial representations of web page documents, search queries, and their subtopics are calculated, involving the following steps:

[0098] (1-1) Input the set of web page documents P = [p1, p2, p3, ..., p N ], where p i Let N represent the i-th document and N represent the total number of documents.

[0099] (1-2) The PV-DBOW training mechanism of the Doc2Vec model is used to obtain document vectors containing word semantics and word order information. In the PV-DBOW training method, the i-th document represents p. i =[w i,1 ,w i,2 ,w i,3 ,...,w i,m ], where w i,j Let represent the j-th word of the i-th document, and m represent the total number of words in the document. Finally, given the text window size, a word is randomly sampled from the text window. By predicting the probability of this word, the document matrix representation is iteratively trained to obtain the document information representation. The goal of the training process is to maximize the probability of the predicted word. The calculation process is as follows:

[0100]

[0101] Where f(p) represents the calculation of the prediction probability of randomly sampled words in document p, and ω (v,z) Let P(ω) represent the z-th randomly sampled predicted word in the v-th text window, W be the set of text windows, v be a randomly sampled text window in W, V be the set of words in window v, and z be a randomly sampled word in V. (v,z) |d) represents generating predicted words ω for a given document p. (v,z) The normalized probability is calculated as follows:

[0102]

[0103] where y k represents the probability of predicting the kth word before normalization, and k k exp(y N ) represents the sum of the probabilities of all predicted words before normalization, and the final document set representation is obtained by updating the document vector through gradient descent method. The information representation calculation formula of the document set is as follows:

[0104] D=Doc2Vec(P)

[0105] where D represents the final document set representation, Doc2Vec(·) represents the Doc2Vec model, and P represents the web document set.

[0106] (1-3) Through the information representation learning of the data set, the web document set P=[p1,p2,p3,...,p N ] is continuously iterated and updated to learn the final document set information representation D=[d1,d2,d3,…,d N ]. When searching for the information representation of the search statement and its subtopics, the relevant search method is used to recall the documents first, and then the top documents are spliced into a new document. The information representation result of the new document is used as the information representation vector of the corresponding search statement or subtopic. In order to be in the same semantic space as the document representation, the search statement and the subtopic are also represented using the Doc2Vec model, and the information representation q of the search statement and the information representation S=[s1,s2,s3,…s k ] of the subtopic are obtained.

[0107] 2. Global diversity representation calculation

[0108] Based on the initial information representation of the web documents, search statements and subtopics obtained in step (1), the self-attention mechanism is used to calculate the novelty representation of the documents and subtopics from the text perspective and the subtopic perspective respectively, and the cross-attention mechanism is used to calculate the cross-attention representation of the mutual fusion of the documents and subtopics. Then, the cross-attention representation of the web documents is used as the query vector, and the cross-attention representation of the subtopics is used as the key vector and the value vector to calculate the coverage representation of the web documents to the subtopics. Finally, the novelty representation of the web documents and the coverage representation of the web documents to the subtopics are combined as the global diversity representation:

[0109] (2-1) Novelty encoding stage, the last layer representation D l-1 =[d1,d2,d3,…,d N ] of the document set is used as the initial input of the self-attention mechanism, and the multi-head self-attention mechanism is used to update the representation between the documents to obtain the document representation D l of the current layer. After L layers of multi-head self-attention calculation, the document self-attention encoding matrix in The vector representation of the i-th document after processing by the novelty encoder is shown. The novelty representation of the multi-layer multi-head self-attention encoder is divided into three parts.

[0110] First, calculate the attention coefficient a. l,t The calculation process is as follows:

[0111] a l,t =Attn(D l-1 D l-1 D l-1 )

[0112] Among them, a l,t This represents the attention coefficient of the t-th group in the l-th layer, where query Q, key K, and value V are all document matrix representations D output by the multi-head self-attention mechanism of the previous layer. l-1 D l-1 The output of the (l-1)th hidden layer is concatenated with h sets of self-attention mechanism coefficients to obtain the multi-head self-attention coefficients (MHSA). l =[a l,1 ;a l,2 ;a l,3 ;…;a l,h ];

[0113] Secondly, the multi-head self-attention representation D of the l-th layer is calculated through a feedforward neural network, residual connections, and layer normalization operations. l =MultiSelfAttn l (D l-1 The calculation process is as follows:

[0114] X = LayerNorm(D) l-1 +MHSA l )

[0115] D l =LayerNorm(X+FeedForward(X))

[0116] Where X represents the input, MHSA l represents the multi-head self-attention coefficient, FeedForward(X) represents the feedforward neural network, and LayerNorm(·) represents the layer normalization operation;

[0117] Finally, a document matrix representation containing novelty information is learned through a multi-layer multi-head self-attention encoder. in This represents the final vector representation of the i-th document after it has been processed by the novelty encoder; its calculation process is as follows:

[0118] D enc= MultiSelfAttn L (MultiSelfAttn L-1 (…MultiSelfAttn1(D)))#

[0119] where MultiSelfAttn L represents the multi-head attention of the L-th layer;

[0120] The calculation method of subtopic novelty representation is consistent with the above method, the only difference is that the initial input is changed to subtopic set information representation S = [s1, s2, s3, … s k , and finally the subtopic novelty representation after multi-layer multi-head self-attention encoding is obtained

[0121]

[0122] (2-2) Subtopic Coverage Decoding Stage, the document novelty representation and the subtopic novelty representation obtained in step (2-1) are taken as input, and the cross-attention between the subtopic novelty representation and the document novelty representation is calculated to calculate the association degree between the document and the subtopic, and the document cross-attention feature representation and the subtopic cross-attention feature representation The calculation process is as follows:

[0123] D co = CrossAttention(D enc ,S enc )

[0124] S co = CrossAttention(S enc ,D enc )

[0125] where CrossAttention(·) represents cross-attention, D enc represents the document novelty representation, and S enc represents the subtopic novelty representation;

[0126] Taking the first layer calculation as an example, the specific calculation of cross-attention is as follows:

[0127] a t = Attention(D co ,S co ,S co )t∈[1,h]

[0128]

[0129] where a tMulti-Head Attention Coefficients of the t-th group of attention heads, and the h Multi-Head Attention Coefficients are spliced to obtain the Multi-Head Attention Coefficients MHA l = [a1; a2; a3;... ; a h ], similar to the encoder in (2-1), the decoder obtains the multi-head attention representation of the document by calculating the multi-head attention coefficients, layer normalization operation, residual connection and feedforward network, and the activation function is ReLU; then, the multi-angle and multi-level attention representation of the document pair sub-topic is obtained by the L-layer multi-head attention mechanism decoder, and the formalized representation is as follows:

[0130] D dec = MultiAttn L (MultiAttn L-1 (…MultiAttn1(D co )))

[0131] D dec is the final representation output by the decoder, representing the coverage degree representation of the document pair sub-topic;

[0132] (2-3) The global diversity representation of the document is obtained by splicing the novelty representation of the document obtained in step (2-1) and the coverage degree representation of the document pair sub-topic obtained in step (2-2);

[0133] 3. Local diversity representation calculation

[0134] In combination with the initial representation of the web document obtained in step (1), the time sequence information representation of the document is calculated using the long short-term memory network according to the document selection order in the selected web document, the time sequence information representation is taken as the node representation of the document graph, and the time sequence correlation representation between the documents is calculated as the local diversity representation by the graph attention mechanism:

[0135] (3-1) The time sequence relationship representation containing the content of the previous document is learned by the LSTM model, and the hidden state representation of the i-th document can be calculated in the hidden layer of the LSTM The calculation process is as follows:

[0136]

[0137] wherein d i represents the initial input representation of the i-th document in the LSTM, represents the time sequence feature representation output by the i-th document after being calculated by the LSTM;

[0138] (3-2) Each document in the set is taken as a node of the graph to construct a document graph, and the edge relationship of the graph is defined by the adjacency matrix E. The adjacency matrix E is an N*N matrix, and e abrepresents the connection relationship between document node a and document node b, if e ab = 1 represents that d a is connected with d b , e ab = 0 represents that d a is not connected with d b ; its definition is as follows:

[0139]

[0140]

[0141] Considering that all the documents in the set are related to the current search, each document learns potential diversity features by relying on the relationship with all other documents, so the document graph is set to a fully connected graph, and the document graph is reconstructed every time the document selection is performed, and the latest time sequence features calculated in step (3-1) are taken as the node representation;

[0142] (3-3) updating the document node using the graph attention network, the document d i updates its own feature information and the feature information of its

[0143] neighbor nodes by aggregation to obtain the feature representation of the current layer The calculation process is as follows:

[0144]

[0145] Wherein, represents the updated representation of the i-th document at the l-th layer, N(d i ) represents the set of all neighbor nodes of the document node d i , d g ∈N(d i ), σ(·) is a nonlinear activation function, Aggregation(·) is an aggregation function, the number of layers of the graph attention network is set to be l, and the updated representation of the i-th document at the l-th layer is its local diversity representation;

[0146] 4. Correlation representation calculation

[0147] In combination with the initial representation of the web documents, search statements and subtopics in step (1), the initial relevance representation between the documents and subtopics is calculated using the method relying on the documents and search statements and the method relying only on the documents, and the obtained initial relevance representation is processed by a multi-layer perception to obtain an updated relevance representation. Considering the theme intersection and semantic redundancy between the contents of the subtopics, the score obtained by processing the novelty representation of the subtopics in step (2) by the multi-layer perception is multiplied by the updated relevance representation as the final relevance representation as a correlation coefficient.

[0148] (4-1) The initial relevance representation of the documents and subtopics is calculated using the calculation methods relying on the documents and search statements: TF-IDF, BM25 and LMIR and the calculation method relying only on the document content: PageRank, inlinks and outlinks Taking TF-IDF as an example, the initial relevance representation calculation process is as follows:

[0149]

[0150]

[0151] score tf-idf =[score tf,body ,score tf,anchor ,score tf,title ,score tf,URL ,score tf,doc ]

[0152] Wherein, represents the relevance representation of the i-th document and the subtopic set under the current search statement, represents the relevance representation of the i-th document corresponding to the k-th subtopic, score tf-idf represents the representation obtained by the TF-IDF method, and the calculation processes of BM25 and LMIR.DIR are similar to those of the TF-IDF method.

[0153] (4-2) The initial relevance feature D S rel After processing by the multi-layer perception, the document corresponding to each subtopic obtains a relevance score, and a total of k relevance scores are obtained as the updated relevance feature D score , and the calculation process is as follows:

[0154] D score =MLP(D S rel )

[0155] di score = [d i,1 score , d i,2 score , d i,3 score ,..., d i,k score ] (4-1)

[0156] where MLP(·) denotes a multi-layer perceptron, d i,k score denotes the relevance score of the k-th subtopic in the i-th document;

[0157] (4-3) The subtopic novelty representation S enc Through multi-layer perceptron processing, each subtopic corresponding to the search statement obtains a novelty value score i , as a weight coefficient of the relevance feature, the weight coefficients of the k subtopics are represented as S w = [score1, score2, score3,..., score k ], the weight coefficient matrix is multiplied by the relevance feature, and the final relevance feature representation D rel is obtained, the calculation process is as follows:

[0158]

[0159]

[0160] where denotes the relevance feature of the i-th document and the k-th subtopic, denotes the relevance feature of the i-th document, D rel denotes the relevance feature matrix representation of all candidate documents;

[0161] 5. Search result diversification sorting module

[0162] The global diversity representation obtained in step (2), the local diversity representation obtained in step (3), and the relevance representation obtained in step (4) are combined to construct a search result diversification module. The global diversity representation, the local diversity representation, and the relevance representation are spliced and the document score is calculated through a multi-layer perceptron. Finally, the top k documents with the highest scores are selected as the most diverse search result list and returned to the user:

[0163] (5-1) The global diversity representation obtained in step (2), the local diversity representation obtained in step (3), and the relevance representation obtained in step (4) are spliced, and the final feature representation of the i-th document is as follows:

[0164]

[0165] wherein, is the final feature representation of the i-th document, d i enc is the novelty representation of the i-th document, d i dec is the subtopic coverage representation of the i-th document, d i GAT is the local diversity representation of the i-th document, d i rel is the relevance feature representation of the i-th document, d

[0166] (5-2) Using a multi-layer perceptron, the score of the document is calculated, and the document is ranked according to the score, and the calculation process is as follows:

[0167] Score = softmax(W o (σ(W F (D feature ))))

[0168]

[0169] wherein, W o ,W F is the parameter matrix of the MLP, D feature is the final matrix representation of the set of all candidate documents, Score is the final score of the document, and the documents are arranged in descending order of the score to obtain the optimal result diversified search ranking list.

[0170] 6, system function display

[0171] The system function display includes data analysis, experimental analysis and result visualization display, wherein the data analysis includes search sentence length distribution, search sentence corresponding subtopic number distribution and search sentence corresponding candidate document number and positive sample number display; the experimental analysis includes comparative experiment and ablation experiment result graph display of the algorithm on the public data set TREC Web Track, and experimental parameter influence analysis graph display of performance; the visualization display includes result diversified search system display, model evaluation index result display and attention visualization analysis display.

[0172] The result diversified method based on multi-view feature fusion of the application runs a result diversified search system based on multi-view fusion, which includes a data processing module, an offline training module and an online search module.

[0173] Among them, the data processing module is responsible for data acquisition and storage work, data acquisition consists of two parts, including the acquisition of TREC Web Track 2009-2012 diversity evaluation data and ClueWeb09 web document data, data storage is responsible for cleaning and storing the data obtained;

[0174] The offline training module uses the data in the data acquisition module to train the model parameters, which is divided into four sub-modules, namely global diversity modeling module, local diversity modeling module, relevance modeling module and result diversification search sorting module. The global diversity modeling module calculates the novelty representation between documents and the coverage representation of subtopics from the global perspective; the local diversity modeling module calculates the time sequence correlation representation between the selected documents and the candidate documents from the local perspective; the relevance modeling module calculates the relevance representation between the search statement and the document; the result diversification search sorting module scores the documents and selects the high-score documents to form the optimal document list.

[0175] The online search module mainly provides user interaction module and visualization display function, mainly including user interaction module, result visualization module and search analysis module. The user interaction module provides the user with the interactive interface of result diversification search; the result visualization module mainly shows the user the result content of diversification search; the search analysis module mainly visualizes the model results, comparative experiments and ablation experiments proposed in this study.

Claims

1. A result diversification search method based on multi-view fusion, characterized in that, The method comprises the following steps: (1) Information representation Through the context analysis and understanding of information, the initial representation of the web document, search statement and sub-topic is calculated; Step (1) comprises the following specific steps: (1-1) Input a set of web page documents P = [p1, p2, p3,..., pn], where pi represents the i-th document, and n represents the total number of documents. N ], where pi represents the i-th document, and n represents the total number of documents. i ], where pi represents the i-th document, and n represents the total number of documents. (1-2) Through the PV-DBOW training mechanism of the Doc2Vec model, a document vector containing word semantics and word sequence information is obtained, and the i-th document representation p i = [w i,1 ,w i,2 ,w i,3 ,...,w i,m ] in the PV-DBOW training method, wherein w i,j represents the j-th word of the i-th document, and m represents the total number of document words; finally, given a text window size, one word in the text window is randomly sampled, the probability of predicting the word is used to iteratively train the document matrix representation, and a document information representation is obtained; the target in the training process is to maximize the probability of predicting the word, and the calculation process is as follows: where f(p) represents the computation of maximizing the predicted probability of a randomly sampled word in the document p, ω (v,z) represents the zth randomly sampled predicted word in the vth sampled text window, W is a set of text windows, v is a randomly sampled text window in W, V represents a set of words in window v, z is a randomly sampled word in V, P(ω (v,z) |p) represents the generation of a predicted word ω (v,z) the normalized probability, which is calculated as follows: where y k represents the probability of predicting the kth word before normalization, and k exp(y k ) represents the sum of the probabilities of all predicted words before normalization, and the final document set representation is obtained by updating the document vector through gradient descent method. The information representation calculation formula of the document set is as follows: D = Doc2Vec (P) Wherein, D represents the final document set representation, Doc2Vec (·) represents the Doc2Vec model, and P represents the web document set; (1-3) Through the representation learning of information in the data set, the web document set P = [p1, p2, p3, …, p N ] is constantly iteratively updated and learned to obtain the final document set information representation D = [d1, d2, d3, …, d N ], when searching for information representation of search statements and sub-topics, first use the relevance search method to recall documents, then splice the top documents into a new document, and use the information representation result of the new document as the information representation vector of the corresponding search statement or sub-topic. In order to be in the same semantic space as the document representation, the search statement and the sub-topic are also represented using the Doc2Vec model to obtain the information representation q of the search statement and the information representation S = [s1, s2, s3, … s k ]. (2) Global diversity representation calculation Combined with the initial information representation of the web document, search statement and sub-topic obtained in step (1), the novelty representation of the document and the sub-topic is calculated from the text perspective and the sub-topic perspective respectively using the self-attention mechanism, and the cross-attention mechanism is used to calculate the cross-attention representation of the document and the sub-topic fusion, then the cross-attention representation of the web document is used as the query vector, the sub-topic cross-attention representation is used as the key vector and the value vector to calculate the coverage representation of the web document to the sub-topic, and finally the novelty representation of the web document and the coverage representation of the web document to the sub-topic are combined as the global diversity representation; Step (2) comprises the following specific steps: (2-1) Novelty encoding stage, the document set on a layer of representation D l-1 = [d1, d2, d3, …, d N ] as the initial input of self-attention mechanism, using multi-head self-attention mechanism to update the representation between documents to get the document representation D l , after L layer multi-head self-attention calculation, finally get the document self-attention encoding matrix Where represents the vector representation of the i-th document after the novelty encoder, the multi-layer multi-head attention encoder divides the novelty representation into three parts, First, the attention coefficient a is calculated l,t The calculation process is as follows: a l,t = Attn(D l-1 , D l-1 , D l-1 ) wherein a l,t represents the t-th group of attention coefficients of the l-th layer, wherein the query Q, the key K and the value V are all document matrix representations D output by the previous layer of multi-head self-attention mechanism l-1 , D l-1 represents the output of the l-1-th hidden layer, and is obtained by splicing the h groups of self-attention mechanism coefficients to obtain the multi-head self-attention coefficient MHSA l = [a l,1 ; a l,2 ; a l,3 ; …; a l,h ]; Second, the multi-head self-attention representation D of the l-th layer is calculated by a feedforward neural network, a residual connection, and a layer normalization operation l = MultiSelfAttn l (D l-1 ), the calculation process is as follows: X = LayerNorm(D l-1 + MHSA l ) D l = LayerNorm(X + FeedForward(X)) wherein X represents an input, MHSA l denotes a multi-head self-attention coefficient, FeedForward(X) denotes a feedforward neural network, and LayerNorm(·) denotes a layer normalization operation. Finally, the document matrix representation containing novelty information is learned by the multi-layer multi-head self-attention encoder wherein represents the final vector representation of the i-th document after the novelty encoder; the calculation process is as follows: D enc = MultiSelfAttn L (MultiSelfAttn L-1 (... MultiSelfAttn1(D))) where MultiSelfAttn L denotes the L-th layer multi-head attention; The calculation manner of the subtopic novelty representation is consistent with the above method, and the only difference is that the initial input is changed to the subtopic set information representation S=[s1, s2, s3, …, s k ], and finally the subtopic novelty representation after multi-layer multi-head self-attention encoding is obtained (2-2) The sub-topic coverage decoding stage takes the document novelty representation and the sub-topic novelty representation obtained in step (2-1) as input, performs cross-attention between the sub-topic novelty representation and the document novelty representation to calculate the association degree between the document and the sub-topic, and obtains a document cross-attention feature representation The sub-topic cross-attention feature representation The calculation process is as follows: D co = CrossAttention(D enc , D enc ) S co = CrossAttention(S enc , D enc ) where CrossAttention(·) denotes cross-attention, D enc denotes the document novelty representation, S enc denotes the subtopic novelty representation; Taking the first layer calculation as an example, the specific calculation of cross-attention is as follows: a t = Attention(D co , S co , S co )t e [1, h] wherein a t denotes the attention coefficient of the t-th attention head, and the h attention coefficients are spliced to obtain the multi-head attention coefficient MHA l = [a1; a2; a3;...; a h ], the decoder obtains the multi-head attention representation of the document by calculating the multi-head attention coefficient, layer normalization operation, residual connection and feedforward network, and the activation function is ReLU; subsequently, the document is obtained by the L-layer multi-head attention mechanism decoder, and the multi-angle and multi-level attention representation of the sub-topic is as follows: D dec = MultiAttn L (MultiAttn L-1 (... MultiAttn1(D co ))) D dec The final representation output by the decoder represents a characterization of the extent to which the document covers subtopics; (2-3) The global diversity representation of the document is obtained by splicing the document novelty representation obtained in step (2-1) and the document coverage representation to the sub-topic obtained in step (2-2); (3) Local diversity representation calculation Combined with the initial representation of the web document obtained in step (1), the time sequence information representation of the document is calculated using the long short-term memory network according to the document selection order in the selected web document, the time sequence information representation is taken as the node representation of the document graph, and the time sequence correlation representation between the documents is calculated as the local diversity representation through the graph attention mechanism; Step (3) comprises the following specific steps: (3-1) The time sequence relationship representation containing the pre-document content is learned by the LSTM model, and the hidden state representation of the i-th document can be calculated in the hidden layer of the LSTM The calculation process is as follows: where d i represents the initial input representation of the i-th document in the LSTM, represents the time-series feature representation output by the LSTM after computing the i-th document; (3-2) Construct a document graph with each document in the set as a node of the graph, and the edge relationship of the graph is defined by an adjacency matrix E, which is an N*N matrix, e ab represents the connection relationship between document node a and document node b, if e ab = 1 indicates that d a is connected with d b , e ab = 0 indicates that d a is not connected with d b ; its definition is as follows: Considering that the documents in the set are all related to the current search, each document learns the potential diversity features by relying on the relationship with all other documents, therefore the document graph is set as a fully connected graph, and the document graph is reconstructed every time the document is selected, and the latest time sequence features calculated in step (3-1) are taken as the node representation; (3-3) Update the document nodes using the graph attention network, document d i The feature representation of the current layer is obtained by aggregating the feature information of the previous layer itself and its neighbor nodes The calculation process is as follows: wherein, represents the updated representation of the ith document at the lth layer, N(d i ) represents the set of all neighbor nodes of the document node d i , d g ∈ N(d i ), σ(·) is a nonlinear activation function, Aggregation(·) is an aggregation function, the number of layers of the graph attention network is set as l, and the updated representation of the ith document at the lth layer is , which is the local diversity representation thereof. (4) Correlation representation calculation Combined with the initial representation of the web document, search statement and sub-topic in step (1), the initial correlation representation between the document and the sub-topic is calculated using the method of relying on the document and the search statement and the method of only relying on the document, and the initial correlation representation obtained is processed through the multilayer perception machine to obtain the updated correlation representation, considering that there is theme intersection and semantic redundancy between the sub-topics, the score of the sub-topic novelty representation calculated in step (2) is multiplied by the updated correlation representation as the final correlation representation through the multilayer perception machine as the correlation coefficient; (5) Search result diversification sorting module The global diversity representation obtained in step (2), the local diversity representation obtained in step (3) and the relevance representation obtained in step (4) are combined to construct a search result diversification module, the global diversity representation, the local diversity representation and the relevance representation are spliced and the document score is calculated through a multi-layer perception, and finally the top k documents with the highest scores are selected as the most diverse search result list and returned to the user. 2.The multi-view fusion based result diversification search method of claim 1, wherein, Step (4) comprises the following specific steps: (4-1) Initial relevance representation of documents to subtopics is computed using document-dependent and query-dependent methods: TF-IDF, BM25 and LMIR, and document-dependent only methods: PageRank, inlinks and outlinks (4-2) The initial relevance feature D obtained in step (4-1) is updated to D S rel After processing by the multi-layer perception, the document obtains a relevance score for each subtopic, and a total of k relevance scores are obtained as the updated relevance feature D score The calculation process is as follows: D score = MLP(D S rel ) d i score = [d i,1 score , d i,2 score , d i,3 score ,..., d i,k score ] where MLP(·) denotes a multi-layer perceptron, d i,k score represents the relevance score of the k-th subtopic in the i-th document. (4-3) The sub-topic novelty representation S obtained in step (2-1) is updated as follows enc Through the multi-layer perception processing, each sub-topic corresponding to the search sentence obtains a novelty value score i , and the weight coefficient of the k sub-topics obtained as the relevance feature is expressed as S w =[score1, score2, score3, …, score k ], the weight coefficient matrix is multiplied by the relevance feature point product, and the final relevance feature representation D rel is obtained; the calculation process is as follows: wherein, represents the relevance feature of the ith document to the kth subtopic, represents the relevance feature of the ith document, D rel represents the relevance feature matrix representation of all candidate documents. 3.The multi-view fusion based result diversification search method of claim 2, wherein, The initial relevance representation calculation process of the calculation method TF-IDF dependent on the document and the search statement in step (4-1) is as follows: score tf-idf = [score tf,body , score tf,anchor , score tf,title , score tf,URL , score tf,doc ] wherein, represents the relevance representation of the i-th document to the subtopic set under the current search statement, represents the relevance representation of the i-th document to the k-th subtopic, Score tf-idf represents the representation obtained by the TF-IDF method. 4.The multi-view fusion based result diversification search method of claim 1, wherein, Step (5) comprises the following specific steps: (5-1) The global diversity representation obtained in step (2), the local diversity representation obtained in step (3) and the relevance representation obtained in step (4) are spliced, and the final feature representation of the i-th document is as follows: wherein, is the final feature representation for the i-th document, d i enc is the novelty representation for the i-th document, d i dec is the subtopic coverage representation for the i-th document, d i GAT is the local diversity representation for the i-th document, d i rel is the relevance feature representation for the i-th document; (5-2) The document score is calculated using a multi-layer perception, and the documents are ranked according to the score, and the calculation process is as follows: Score = softmax(W o (σ(W F (D feature )))) Among them, W o W F Let D be the parameter matrix of the MLP. feture The final matrix representation of the set of all candidate documents is represented by the Score, which is the final score of each document. The documents are then sorted from highest to lowest score to obtain the optimal result: a diversified search ranking list.

5. A result diversification system based on multi-view feature fusion, characterized in that, The system is used to run the result diversification search method based on multi-view fusion according to any one of claims 1-4, and the system comprises a data processing module, an offline training module and an online search module; The data processing module is responsible for data acquisition and storage, and the data acquisition comprises two parts, including acquiring the diversification evaluation data of TREC Web Track 2009-2012 and the webpage document data of ClueWeb09, and the data storage is responsible for cleaning and storing the acquired data; The offline training module uses the data in the data acquisition module to train the model parameters, and the module comprises four sub-modules, namely a global diversity modeling module, a local diversity modeling module, a relevance modeling module and a result diversification search ranking module; the global diversity modeling module calculates the novelty representation between documents and the coverage degree representation of a document to a sub-topic from a global perspective; the local diversity modeling module calculates the time sequence correlation representation between the selected documents and the candidate documents from a local perspective; the relevance modeling module calculates the relevance representation between the search statement and the document; and the result diversification search ranking module scores the documents and selects the documents with high scores to form an optimal document list in combination with the above representations; The online search module mainly provides a user interaction module and a visual display function, and comprises a user interaction module, a result visualization module and a search analysis module; the user interaction module provides an interactive interface for the result diversification search for the user; the result visualization module mainly displays the result content of the diversification search for the user; and the search analysis module mainly visualizes the model results, the comparative experiments and the ablation experiments of the present research.

Images