A view retrieval system based on two-path encoding and exact match signals

Abstract

The application provides a kind of view retrieval system based on double-path encoding and accurate matching signal, comprising: a query text and document text preprocessing module, and the input query and candidate document are preprocessed;A query text and document text relevance score calculation module;A query text and document text view score calculation module, the view score of candidate document is calculated by pre-training model;Unified correlation retrieval module is used to obtain the view score of document according to the relevance score of query and document obtained by correlation retrieval module and according to view score module, finally calculate the view retrieval score of document.This technical solution obtains local semantic information and global semantic information by double-path encoding, can carry out view retrieval by fusing and based on these information, and the accurate semantic information obtained by accurate matching mechanism can improve the relevance of query and document.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of opinion retrieval, and particularly relates to an opinion retrieval system based on double coding and accurate matching signals. BACKGROUND

[0002] With the rapid development of Web2.0 and the popularity of the Internet, a large number of social-based platforms (such as Zhihu, Sina Weibo, Bilibili, etc.) appear on the Internet. More and more people are keen on publishing and sharing their views on hot information on social media, and these social platforms have become a carrier for people to express their opinions. By analyzing these opinion information, it can be understood. Opinion mining aiming at sentiment analysis and information mining (mining opinion information of text) has become one of the frontiers of the field of natural language processing. Network text opinion retrieval is to study how to retrieve the documents related to the query text and having a certain subjective tendency from a large number of social text documents. The research and discussion on the topic of text opinion retrieval have attracted widespread attention from scholars in academia, industry and various industries. The research of text opinion retrieval model has experienced three stages: the early two-stage retrieval model, the linear retrieval model, and finally the unified relevance model. The three stages will be briefly introduced below.

[0003] The two-stage retrieval model first retrieves the documents related to the given query topic, then identifies the tendency of these documents related to the given query, and finally sorts the documents according to the relevance and tendency. This model has a simple structure and is easy to understand, but lacks a reasonable theoretical explanation. The first stage of this method usually uses classical retrieval models such as language models and BM25, while the research focus is on the second stage, i.e. opinion mining of documents.

[0004] Linear retrieval models provide researchers with a new way of thinking, i.e. using a final indicator to score the documents, which also promotes the research of unified opinion retrieval models. Unlike the traditional document representation method of topic-opinion word representation method, the context information between the opinion word in the sentence and its target is captured, and the relationship between multiple opinion word sentences of the same topic is considered, and the two kinds of information are combined into a unified graph model. The HITS algorithm is used to calculate the document score and sort.

[0005] Unified retrieval model, with the help of the latest models in the fields of current information retrieval and text mining, directly mines the tendency of the description topic to sort the documents. Compared with the two-stage model, this method has the advantages of easy interpretation in theory, more direct and effective expression of information needs, etc. SUMMARY

[0006] Therefore, the present application aims to provide a viewpoint retrieval system based on double encoding and accurate matching signals, which can obtain local semantic information and global semantic information through double encoding, can perform viewpoint retrieval by fusing and based on the information, and can improve the relevance of queries and documents through accurate semantic information obtained through an accurate matching mechanism.

[0007] To achieve the above object, the present application adopts the following technical solution: a viewpoint retrieval system based on double encoding and accurate matching signals, comprising:

[0008] A query text and document text preprocessing module is used to preprocess input queries and candidate documents, mainly removing irrelevant information such as symbols, emoticons, and stop words, and performing word segmentation, encoding, and semantic space mapping to obtain word vectors and sentence representation vectors;

[0009] A query text and document text relevance score calculation module is used to obtain sentence encoding through a transformer and a recurrent neural network LSTM, then fuse the two obtained encoding information, add the sentence representation to an accurate matching network to obtain accurate matching semantic information, and finally use a Gaussian kernel pooling method to obtain the relevance score between the query and the document;

[0010] A query text and document text viewpoint score calculation module is used to pre-train a viewpoint feature extraction model to calculate the viewpoint score of the candidate document through the pre-trained model;

[0011] A unified relevance retrieval module is used to obtain the viewpoint score of the document according to the relevance score of the query and the document obtained by the relevance retrieval module and the viewpoint score module, finally calculate the viewpoint retrieval score of the document, sort the document, and output the sorting result.

[0012] In a preferred embodiment: the query text and document text preprocessing module removes symbols, emoticons, network links, and stop words irrelevant information from the query and document data, and performs word segmentation, establishes a word table, encodes, and maps the semantic space to obtain word vectors and sentence representation vectors.

[0013] In a preferred embodiment: the query text and document text relevance score calculation module is composed of a double-context encoding layer, a context fusion layer, an accurate matching layer, a convolution layer, a matching layer, a kernel pooling layer, and a learning layer;

[0014] The double-context encoding layer is composed of a TransformerEncoder and an LSTM recurrent neural network, and is used to obtain context vectors corresponding to global context semantic representation and local context semantic representation according to input word vectors;

[0015] The TransformerEncoder formula is as follows:

[0016] X' = Transformer(X)

[0017] Transformer(X) = LayerNorm(LayerNorm(MultiHeadAttention(LL(X)) + X) + FFN(LayerNorm(MultiHeadAttention(LL(X)) + X)

[0018] LL(X) = Linear(ReLU(Linear(X)))

[0019] MultiHeadAttention(X) = Concat(head1,..., head h )W O

[0020]

[0021] Linear(X) = XA T + b

[0022] FFN(X) = max(0, XW1 + b1)W2 + b

[0023]

[0024] where X is the word vector corresponding to the input, MultiHeadAttention is the multi-head self-attention module, ReLU is the nonlinear activation function, W i is the parameter matrix that needs to be learned by the model;

[0025] The LSTM formula is represented as follows:

[0026] f t = σ(W f · [h t-1 , x t ] + b f )

[0027] i t = σ(W i · [h t-1 , x t ] + b i )

[0028]

[0029]

[0030] Ot = σ(W o [h t-1 , x t ]+ b o )

[0031] h t = O t * tanh(C t )

[0032] T i = h t

[0033] where σ denotes the logistic sigmoid function; tanh denotes the Hyperbolic tangent function; W f , W i , W c , W o denote trainable parameter matrices; b f , b i , b c , b o denote corresponding biases; i t denotes the input gate at time t, which decides the information that needs to be preserved for the current time step input; f t denotes the forget gate at time t, which is used to control the proportion of information that should be lost from the historical cell state; o t denotes the output gate at time t, which is used to control the output of the hidden state; denotes the current input feature calculated by the nonlinear function tanh; C t denotes the memory cell state at time t; h t denotes the hidden layer output vector at time t;

[0034] The LSTM structure calculates the current node state according to the previous time state to obtain local context semantic information;

[0035] X'' = Gate(Linear(Concat(Transformer(X), LSTM(X))))

[0036] Gate(Y) = tanh(W * (σ(W1 * (Y, h) + e(W2 * (Y, h))))

[0037] The context fusion layer is mainly composed of a fully connected layer module and a gating network. The global context semantic information and the local context semantic information are fused in the first step by splicing. At this time, the dimension of each representation vector becomes twice as large as the original one. Then, these vectors are reduced in dimension by the first fully connected layer. The dimension of each representation vector is returned to the original dimension to obtain the global and local fused context representation vector. At this time, the context vector corresponding to each word not only contains the information of the word itself, but also carries different levels of rich context information. i is the parameter matrix that the model needs to learn, σ represents the logistic sigmoid function, Y is the concatenated semantic information, h is the hidden state of the previous node, and tanh() is the hyperbolic tangent function;

[0038] The calculation formula of the exact matching layer is:

[0039] Dot product function: The most commonly used calculation method, query vector q and document vector q are directly dot-producted to obtain the relevance score:

[0040] rel1(q, d) = q T d

[0041] Scaled dot product function: query vector q and document vector d perform dot product operation and then multiply by the scaling factor Get the relevance score, where dim is the dimension of the query vector q:

[0042]

[0043] Then pass

[0044] similar(q, d)

[0045] =Linear(Linear(Concat(rel1, rel2)))+Linear(q)

[0046] The convolution layer consists of several convolution kernels with different window sizes. Through these convolution kernels, the adjacent position representation information in the input text can be further obtained based on the fusion of local context information and global context information obtained in the previous layer.

[0047] C h =ReLU(W h ·X″+b h ), i = 1, 2, ..., n

[0048] Where h∈[1, N] is the window size of the convolution kernel in the word dimension, N=2, and W is the parameter matrix of the convolution kernel;

[0049] The matching layer calculates the user query based on the N-Gram information obtained from the previous layer and candidate documents between each N-Gram vector, construct the query-candidate document interaction matrix:

[0050]

[0051] The core pooling layer is composed of k Gaussian kernel functions for generating soft interaction features, which captures the implicit soft interaction features φ(M) by calculating the N-Gram similarity scores of k intensity levels in the interaction matrix through k Gaussian kernel functions:

[0052]

[0053]

[0054]

[0055] The learning layer obtains the final relevance score RelScore by learning the soft interaction features extracted from the previous layer;

[0056] RelScore(q, d) = sigmoid(Linear(φ(M))).

[0057] In a preferred embodiment: the query text and document text opinion score calculation module introduces a pre-trained model, by training a pre-trained model for opinion feature extraction, introducing the Encoder structure of the large corpus pre-trained model Bert, extracting the deep context semantic information of the text, and training the opinion feature extractor conforming to the system through fine-tuning, thereby obtaining the opinion score of the candidate document.

[0058] In a preferred embodiment: the unified relevance retrieval module, by fusing the relevance score of the document and the query with the opinion score of the document through the unified opinion retrieval model, obtains the opinion retrieval score of the document about the query, RankScore(q, d) = RelScore(q, d)·OpiScore(d), and reorders all candidate documents according to the obtained opinion retrieval score of the document, and outputs the final sorted score set result.

[0059] Compared with the prior art, the present application has the following beneficial effects: global semantic information and local semantic information can be obtained through the recurrent neural network LSTM and the multi-head attention transformer, more rich context semantic information of the text data is obtained through the fusion of the two, the precise semantic information of the query is obtained by the precise matching mechanism, the N-Gram information of the text data is obtained by the convolutional neural network, the cosine similarity of the query and the document is calculated, the result is obtained through the kernel pooling method, and the viewpoint retrieval performance is better based on the methods. And a viewpoint feature extraction pre-training model is included, deep viewpoint features are extracted through the multi-layer Bert encoder method, and the viewpoint retrieval performance is further improved BRIEF DESCRIPTION OF DRAWINGS

[0060] Fig. 1 The model schematic diagram of the preferred embodiment of the present application is shown.

[0061] Fig. 2 The model schematic diagram of the viewpoint score calculation module of the preferred embodiment of the present application is shown. DETAILED DESCRIPTION

[0062] The present application will be further described below in combination with the drawings and embodiments.

[0063] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs.

[0064] It should be noted that the terms used herein are only for the purpose of describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application; as used herein, the singular form is intended to include the plural form unless the context clearly indicates otherwise, and furthermore, it should be understood that when the terms "comprise" and / or "include" are used in the specification, there is a feature, step, operation, device, component and / or combination thereof.

[0065] As shown in Figs. 1-2 A viewpoint retrieval system based on double encoding and precise matching signal, comprising:

[0066] The text preprocessing module removes irrelevant information such as symbols, emoticons, network links, stop words, etc. from the query and document data, and establishes a word table, encodes and maps the semantic space to obtain word vectors and sentence representation vectors. The relevance score calculation module obtains sentence encoding by introducing transformer and recurrent neural network LSTM, adds sentence representation to the exact matching network to obtain exact matching semantic information, introduces convolutional neural network to calculate the N-Gram information of the text, obtains relevance information by calculating the cosine similarity of the document and the query, and finally obtains the relevance score between the query and the document by using the Gaussian kernel pool method. The opinion score calculation module introduces a pre-trained model. By training a pre-trained model for opinion feature extraction, the Encoder structure of the large corpus pre-trained model Bert is introduced to extract deep context semantic information of the text. The opinion score calculation module is trained on a short text review dataset through word embedding and multi-layer Enconder structure, and the opinion feature extractor conforming to the system is trained through fine-tuning and feedforward neural network, so as to obtain the opinion score of the candidate document. The unified relevance retrieval module is used to obtain the opinion score of the document according to the relevance score of the query and the document obtained by the relevance retrieval module and the opinion score module, and finally calculate the opinion retrieval score of the document. The documents are sorted according to the opinion retrieval score, and the sorting result is obtained.

[0067] 1) Text processing module 1

[0068] First, how the text preprocessing module preprocesses the text is described.

[0069] The social media document contains rich information but also contains some noise, so the data set is preprocessed. The following operations are mainly performed:

[0070] (1) Remove punctuation, emoticons, network links, stop words and other irrelevant information in the document;

[0071] (2) Tokenize the text data and establish a word table;

[0072] (3) Encode the text data according to the word table and restore the stem;

[0073] (4) Obtain the word vector of the text data according to the pre-trained language model Glove;

[0074] 2) Relevance score calculation module 2

[0075] The following describes how the relevance score calculation module 2 uses the text vector representation obtained by the text preprocessing module 1 to calculate the relevance score of the query text and the candidate document text by fusing double-channel encoding and exact matching signal module.

[0076] Specifically, as shown in Fig. 2 , this module is composed of a two-way context encoding layer, a context fusion layer, an exact matching layer, a convolution layer, a matching layer, a kernel pooling layer, and a learning layer.

[0077] The two-way context encoding layer is composed of a TransformerEncoder and an LSTM recurrent neural network, which is used to obtain a context vector corresponding to a global context semantic representation and a local context semantic representation according to an input word vector.

[0078] The TransformerEncoder formula is as follows:

[0079] X' = Transformer(X)

[0080] Transformer(X) = LayerNorm(LayerNorm(MultiHeadAttention(LL(X)) + X) + FFN(LayerNorm(MultiHeadAttention(LL(X)) + X)

[0081] LL(X) = Linear(ReLI(Linear(X)))

[0082] MultiHeadAttention(X) = Concat(head1,..., head h )W O

[0083]

[0084] Linear(X) = XA T + b

[0085] FFN(X) = max(0, XW1 + b1)W2 + b

[0086]

[0087] where X is the word vector corresponding to the input, MultiHeadAttention is a multi-head self-attention module, ReLU is a nonlinear activation function, and W i is a parameter matrix that needs to be learned by the model.

[0088] The LSTM formula is represented as follows:

[0089] f t = σ(W f ·[h t-1 , x t ] + bf )

[0090] i t = σ(W i · [h t-1 , x t ] + b i )

[0091]

[0092]

[0093] O t = σ(W o [h t-1 , x t ] + b o )

[0094] h t = O t * tanh(C t )

[0095] T i = h t

[0096] where σ denotes the logistic sigmoid function; tanh denotes the Hyperbolic tangent function; W f , W i , W c , W o denote trainable parameter matrices; b f , b i , b c , b o denote the corresponding biases.i t denotes the input gate at time t, which decides the information that needs to be preserved for the current time step input; f t denotes the forget gate at time t, which is used to control the proportion of information that should be lost from the historical cell state; o t denotes the output gate at time t, which is used to control the output of the hidden state; denotes the current input feature calculated by the nonlinear function tanh; C t denotes the memory cell state at time t; h t denotes the hidden layer output vector at time t.

[0097] The LSTM structure calculates the current node state according to the previous time step state to obtain local context semantic information.

[0098] X" = Gate(Linear(Concat(Transformer(X), LSTM(X))))

[0099] Gate(Y) = tanh(W*(σ(W1*(Y, h) + σ(W2*(Y, h))

[0100] The context fusion layer is mainly composed of a full connection layer module and a gate network. Global context semantic information and local context semantic information are fused in a first step by splicing, so that the dimension of each representation vector is doubled. Then, the vectors are reduced in dimension by a first full connection layer, and the dimension of each representation vector is returned to the original dimension to obtain a global and local fused context representation vector. At this time, the context vector corresponding to each word contains not only the information of the word itself, but also rich context information at different levels. i is a parameter matrix to be learned by the model, σ represents a logistic sigmoid function, Y is the spliced semantic information, h is the hidden state of the previous node, tanh() is a hyperbolic tangent function.

[0101] The relevant calculation formula of the exact match layer is:

[0102] Dot product function: the most commonly used calculation method, the query vector q and the document vector d are directly dot multiplied to obtain the relevance score:

[0103] rel1(q, d) = q T d

[0104] Scaled dot product function: the query vector q and the document vector d are dot multiplied and then multiplied by a scaling factor to obtain the relevance score, where dim is the dimension of the query vector q:

[0105]

[0106] Then

[0107] similar(q, d)

[0108] = Linear(Linear(Concat(rel1, rel2))) + Linear(q)

[0109] The convolution layer is composed of a plurality of convolution kernels with different window sizes. Based on the fused local context information and global context information obtained from the previous layer, the convolution kernels can further obtain adjacent position representation information in the input text.

[0110] C h = ReLU(W h · X" + b h ), i = 1, 2,..., n

[0111] Where h∈[1, N] is the window size of the convolution kernel in the word dimension, N=2, and W is the parameter matrix of the convolution kernel.

[0112] The matching layer calculates the user query based on the N-Gram information obtained from the previous layer and candidate documents The similarity score of each N-Gram vector between them is used to construct the query-candidate document interaction matrix:

[0113]

[0114] The kernel pooling layer consists of k Gaussian kernel functions to generate soft interaction features. This layer calculates the N-Gram similarity scores of k intensity levels in the interaction matrix through k Gaussian kernel functions to capture the implicit soft interaction features φ(M):

[0115]

[0116]

[0117]

[0118] The learning layer obtains the final relevance score RelScore by learning based on the soft interaction features extracted by the previous layer.

[0119] RelScore(q,d)=sigmoid(Linear(φ(M)))

[0120] 3) Opinion score calculation module 3

[0121] The following describes how the opinion score calculation module 3 uses the text word vector representation obtained by the text preprocessing module 1 to calculate the opinion score of the candidate document text through the pre-trained opinion feature extraction module.

[0122] Specifically, if Fig. 2 As shown in Figure 2, this module is composed of a Bert encoder, a context fusion layer, a convolutional layer, and a classifier. The hierarchical context encoding layer, the context fusion layer, and the convolutional layer are the same as those in the correlation score calculation module 2. The classifier consists of two fully connected layers:

[0123]

[0124] 4) Unified related search module 4

[0125] The following describes how the unified relevance retrieval module 4 calculates the final opinion retrieval score RankScore using the relevance score obtained by the relevance score calculation module 2 and the opinion score obtained by the opinion score calculation module 3 .

[0126] The relevant score is multiplied with the query score to obtain the opinion retrieval score of the candidate document:

[0127] RankScore(q, d) = RelScore(q, d) - OpiScore(d)

[0128] Finally, the documents are ranked according to the opinion retrieval scores of each document from high to low, and the document list is output in sequence to obtain the final opinion retrieval result.

Claims

1. An opinion retrieval system based on two-way coding and exact matching signals, characterized in that: include: A query text and document text preprocessing module preprocesses the input query and candidate document data, mainly removing symbols, emoticons, stop words and irrelevant information from the data, and segmenting, encoding and mapping the data into semantic space to obtain word vectors and sentence representation vectors; A module calculates the correlation score between query text and document text. It uses a transformer and a recurrent neural network (LSTM) to obtain sentence encodings. It then fuses the encoding information obtained from the two paths and feeds the sentence representation into an exact matching network to obtain exact matching semantic information. Finally, it uses the Gaussian kernel pooling method to obtain the correlation score between the query and the document. A query text and document text opinion score calculation module, which calculates the opinion scores of candidate documents by pre-training a opinion feature extraction model; A unified relevance retrieval module is used to calculate the opinion retrieval score of the document based on the relevance score of the query and the document obtained by the relevance retrieval module and the opinion score of the document obtained by the opinion score module, sort the documents, and output the sorting results; The query text and document text correlation score calculation module is composed of a dual-path context encoding layer, a context fusion layer, an exact matching layer, a convolution layer, a matching layer, a kernel pooling layer, and a learning layer. The dual-path context encoding layer consists of a TransformerEncoder and an LSTM recurrent neural network, which is used to obtain context vectors corresponding to global context semantic representation and local context semantic representation based on the input word vector; The TransformerEncoder formula is as follows: X′=Transformer(X) Transformer(X)=LayerNorm(LayerNorm(MultiHeadAttention(LL(X))+X)+ FFN(LayerNorm(MultiHeadAttention(LL(X))+X))) LL(X)=Linear(ReLU(Linear(X))) MultiHeadAttention(X)=Concat(h ead1,…,h ead h )W O Linear(X)=XA T +b FFN(X)=max(0,XW1+b1)W2+b Where X is the word vector corresponding to the input, MultiHeadAttention is the multi-head self-attention module, ReLU is the nonlinear activation function, and W i is the parameter matrix that the model needs to learn; The LSTM formula is expressed as follows: f t =σ(W f ·[h t-1 ,x t ]+b f ) i t =σ(W i ·[h t-1 ,x t ]+b i ) The t =σ(W o [h t-1 ,x t ]+b o ) h t =O t *fishy(C) t ) T i =h t Where σ represents the logistic sigmoid function; tanh represents the Hyperbolic tangent function; W f 、W i 、W c 、W o represents the trainable parameter matrix; b f 、b i 、b c 、b o Indicates the corresponding bias; i t Represents the input gate at time t, which determines the information that needs to be retained in the current time step input; f t Represents the forget gate at time t, which is used to control the proportion of information that should be lost in the historical cell state; o t Represents the output gate at time t, which is used to control the output of the hidden state; Represents the current input feature calculated by the nonlinear function tanh; C t represents the state of the memory unit at time t; h t represents the hidden layer output vector at time t; The LSTM structure calculates the current node state based on the state of the previous time step and obtains local context semantic information; X″=Gate(Linear(Concat(Transformer(X),LSTM(X)))) Gate(Y)=tanh(W*(σ(W1*(Y,h)+σ(W2*(Y,h))))) The context fusion layer is mainly composed of a fully connected layer module and a gating network. The global context semantic information and the local context semantic information are fused in the first step by splicing. At this time, the dimension of each representation vector becomes twice as large as the original one. Then, these vectors are reduced in dimension by the first fully connected layer. The dimension of each representation vector is returned to the original dimension to obtain the global and local fused context representation vector. At this time, the context vector corresponding to each word not only contains the information of the word itself, but also carries rich context information at different levels. i is the parameter matrix that the model needs to learn, σ represents the logistic sigmoid function, Y is the semantic information of the splicing, h is the hidden state of the previous node, and tanh() is the hyperbolic tangent function; The calculation formula of the exact matching layer is: Dot product function: The most commonly used calculation method, query vector q and document vector q are directly dot-producted to obtain the relevance score: rel1(q,d)=q T d Scaled dot product function: query vector q and document vector d perform dot product operation and then multiply by the scaling factor Get the relevance score, where dim is the dimension of the query vector q: Then pass similar(q,d) =Linear(Linear(Concat(rel1,rel2)))+Linear(q) The convolution layer consists of several convolution kernels with different window sizes. Through these convolution kernels, the adjacent position representation information in the input text can be further obtained based on the fusion of local context information and global context information obtained in the previous layer. C h =ReLU(W h ·X”+b h ),i=1,2,…,n Where h∈[1,N] is the window size of the convolution kernel in the word dimension, N=2, and W is the parameter matrix of the convolution kernel; The matching layer calculates the user query based on the N-Gram information obtained from the previous layer and candidate documents The similarity score of each N-Gram vector between them is used to construct the query-candidate document interaction matrix: The kernel pooling layer consists of k Gaussian kernel functions to generate soft interaction features. This layer calculates the N-Gram similarity scores of k intensity levels in the interaction matrix through k Gaussian kernel functions to capture the implicit soft interaction features φ(M): The learning layer obtains the final relevance score RelScore by learning the soft interaction features extracted by the previous layer; RelScore(q,d)=sigmoid(Linear(φ(M))); The query text and document text opinion score calculation module introduces a pre-trained model. By training a pre-trained model for opinion feature extraction and introducing the Encoder structure of the large corpus pre-training model Bert, the deep contextual semantic information of the text is extracted. After fine-tuning, an opinion feature extractor that meets the requirements of this system is trained to obtain the opinion score of the candidate document.

2. The opinion retrieval system based on dual-path coding and exact signal matching according to claim 1, characterized in that: The query text and document text preprocessing module removes symbols, emoticons, network links, stop words and irrelevant information from the query and document data, and performs word segmentation on the data to establish a vocabulary, encode and map the semantic space to obtain word vectors and sentence representation vectors.

3. The opinion retrieval system based on dual-path coding and exact signal matching according to claim 1, characterized in that: The unified relevance retrieval module obtains the opinion retrieval score of the document regarding the query by fusing the relevance scores of the document and the query with the opinion score of the document through a unified opinion retrieval model, RankScore(q,d)=RelScore(q,d)·OpiScore(d), and re-ranks all candidate documents according to the obtained opinion retrieval score of the document, and outputs the final ranked score set result.

Images