CN112949318B - Text position detection method based on text and user representation learning - Google Patents

Text position detection method based on text and user representation learning Download PDF

Info

Publication number
CN112949318B
CN112949318B CN202110233476.XA CN202110233476A CN112949318B CN 112949318 B CN112949318 B CN 112949318B CN 202110233476 A CN202110233476 A CN 202110233476A CN 112949318 B CN112949318 B CN 112949318B
Authority
CN
China
Prior art keywords
vector
text
user
matrix
layer
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202110233476.XA
Other languages
Chinese (zh)
Other versions
CN112949318A (en
Inventor
彭愈翔
罗绪成
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
University of Electronic Science and Technology of China
Original Assignee
University of Electronic Science and Technology of China
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by University of Electronic Science and Technology of China filed Critical University of Electronic Science and Technology of China
Priority to CN202110233476.XA priority Critical patent/CN112949318B/en
Publication of CN112949318A publication Critical patent/CN112949318A/en
Application granted granted Critical
Publication of CN112949318B publication Critical patent/CN112949318B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/30—Semantic analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/047—Probabilistic or stochastic networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent

Abstract

The invention discloses a text standpoint detection method based on text and user representation learning, which comprises the steps of obtaining a text data set from a social media platform, generating a user social relationship diagram and obtaining a corresponding Laplacian matrix, determining a standpoint label vector of each text, obtaining a text vector of each text by adopting a pre-trained BERT model, constructing and training a standpoint detection model, generating the user social relationship diagram and obtaining the corresponding Laplacian matrix when the text of a user needs to be subjected to the standpoint detection, and inputting the text vector into the standpoint detection model to obtain a standpoint detection result. The invention respectively obtains two modal characteristics of the text and the user and carries out cross-modal fusion, thereby realizing accurate text detection.

Description

Text position detection method based on text and user representation learning
Technical Field
The invention belongs to the technical field of natural language processing, and particularly relates to a text position detection method based on text and user representation learning.
Background
The position detection is one of the leading research branches in the field of Natural Language Processing (NLP), and aims to automatically detect the opinion or attitude, such as support, objection or neutrality, expressed by a person on an individual, thing or event from text information.
At present, the conventional vertical detection method mainly adopts a classical model mainly including a CNN (Convolutional Neural Networks) model and an RNN (Recurrent Neural Networks) model. Most of the existing position detection methods only use the information of text dimension for position detection, and do not utilize the user characteristics related to the position height. The attributes of the users and the social relations among the users can influence the places where the users make the statements. Secondly, both CNN and RNN models have their limitations on NLP tasks, for example, CNN can only see local areas, RNN has more attenuation of information as the distance increases for the forward input due to its structure. In recent years, with the proposal of a pre-training language model BERT, the performance of each subtask of NLP is greatly improved. Better vector representation of text dimensions can be obtained by the BERT model. Finally, since the information dimensions of the user and the text are different, the user can be regarded as different modalities. Different modalities have different semantics, and simple splicing cannot realize effective mixing of the semantics among the modalities.
Disclosure of Invention
The invention aims to overcome the defects of the prior art and provides a text position detection method based on text and user representation learning, which respectively obtains two modal characteristics of a text and a user and performs cross-modal fusion, thereby realizing accurate text position detection.
In order to achieve the above object, the text position detection method based on text and user representation learning of the present invention comprises the following steps:
s1: determining a social media platform needing text position detection, collecting a text data set of a topic needing text position detection from the social media platform, wherein the text data set comprises a plurality of texts related to the topic, and an attention list and an attention-to-be-paid list among users publishing the texts;
generating a directed acyclic user social relationship graph G according to the attention list and the attention list among the users<V,E>Wherein V represents all user ID sets, E represents a set of directed edges between users, if a user i pays attention to a user j, a directed edge from the user i to the user j exists, and i, j belongs to V; and then constructing a user social relationship graph G ═<V,E>Of a neighboring matrixA and a degree matrix D; performing Laplace matrix transformation on the adjacent matrix A and the degree matrix D to obtain a Laplace matrix Lrw；
Standardizing the position of a user in a text into three values of objection, neutrality and approval, coding in a one-hot form to obtain a 3-dimensional position vector, and determining the position vector corresponding to each text as a position label vector according to the position of each text;
s2: unifying texts in the text data set into a preset length W: if the text length is larger than W, deleting the excess text, and if the text length is smaller than W, filling preset characters; respectively inputting each text obtained by processing into a pre-trained BERT model, and taking the output d-dimensional vector as a text vector vtWherein t represents a text, t belongs to phi, phi represents a text set in the text data set, and the size of d is set according to actual needs;
s3: constructing a vertical detection model, which comprises a GCN network, an interaction layer, a decision layer and a full connection layer, wherein:
the GCN is used for generating user vectors, and the specific method comprises the following steps: taking a Laplacian matrix of a user social relationship graph as an adjacency matrix of a GCN (GCN network), processing an initial vector matrix of a user by the GCN, taking an obtained output matrix as a user vector matrix, and taking each row of vectors as user vectors corresponding to the user;
the interaction layer is used for performing semantic fusion on the text vector and a user vector of a user to which the text obtained by the GCN belongs to obtain an interaction vector; the interaction layer comprises a hidden layer and an attention layer, wherein the hidden layer is used for solving the outer product of the text vector and the user vector to obtain a dxd interaction matrix; the attention layer is used for reducing the dimension of the interaction matrix to obtain a d-dimension interaction vector;
the decision layer is used for carrying out weighted summation on the user vector, the text vector and the interaction vector to obtain a weighted synthetic vector; the decision layer comprises a splicing layer, an MLP (Multi-level hierarchical processing) network, a softmax layer and a weighting synthesis module, wherein the splicing layer is used for splicing the user vector, the text vector and the interaction vector to obtain a 3 xd splicing matrix; MLP network for obtaining implicit expression vector u of each row vector in splicing matrixj，j＝1,2,3；soThe ftmax layer is used to represent the vector u according to 3 implicit representationsjDetermining a normalized weight αjThe calculation formula is as follows:
Figure BDA0002959626790000021
wherein u iswIs a d-dimensional vector for implicitly representing a vector ujDimension reduction is a numerical value;
a weighted synthesis module for synthesizing the weight alphajAnd carrying out weighted summation on the user vector, the text vector and the interaction vector to obtain a weighted composite vector s, wherein the weighted summation formula is as follows:
s＝∑αjhj
wherein h is1,h2,h3Respectively representing a user vector, a text vector and an interaction vector;
the full-connection layer is used for processing the weighted synthesis vector s to obtain a 3-dimensional vector, wherein each element represents the probability that the text belongs to three positions of objection, neutrality and praise;
s4: the laplace matrix L in step S1 is usedrwAs an adjacency matrix of a GCN (GCN network) in the vertical detection model, randomly generating a d-dimensional initial vector of each user in the text data set and forming a user initial vector matrix U as a row vector, taking the user initial vector matrix U and the text vector obtained in the step S2 as input, taking a vertical label vector corresponding to the text vector as expected output, and training the vertical detection model;
s5: when text published by users of the social media platform needs to be detected from the standpoint, firstly, an attention list and an attention list between current users of the social media platform are obtained, a directed acyclic user social relationship graph G ' is generated, an adjacency matrix A ' and a degree matrix D ' are constructed, and laplacian matrix transformation is performed by adopting the same method of the step S1 to obtain a laplacian matrix Lrw′(ii) a Obtaining a d-dimensional text vector v' from the text needing to be subjected to the position detection by adopting the same method in the step S2;
will Laplace matrix Lrw′And as an adjacency matrix of a GCN (GCN network) in the position detection model, randomly generating d-dimensional initial vectors of all users in a user social relationship graph G ', forming a user initial vector matrix U' as row vectors, and inputting the user initial vector matrix U 'and the text vector v' into the position detection model to obtain a position detection result.
The text position detection method based on text and user representation learning comprises the steps of obtaining a text data set from a social media platform, generating a user social relationship diagram and obtaining a corresponding Laplace matrix, determining a position label vector of each text, obtaining a text vector of each text by adopting a pre-trained BERT model, constructing and training a position detection model, generating the user social relationship diagram and obtaining the corresponding Laplace matrix when position detection needs to be carried out on the user text, and inputting the text vector into the position detection model to obtain a position detection result.
The invention has the following beneficial effects:
1) the internal relations among different users are fully utilized to model the users in the user dimension, and the learned user vector contains the social topological relations among the users and the closeness degree of the relations, so that the method has stronger feature expression capability;
2) the text vector and the user vector are subjected to cross-modal fusion in the position detection model, so that the effective mixing of semantics between the two modalities of the user and the text is realized, the text position detection aiming at the user is realized, and the efficiency and the accuracy of the position detection are improved.
Drawings
FIG. 1 is a flow diagram of an embodiment of a text standpoint detection method based on text and user representation learning of the present invention;
FIG. 2 is a structural diagram of a vertical inspection model according to the present invention.
Detailed Description
The following description of the embodiments of the present invention is provided in order to better understand the present invention for those skilled in the art with reference to the accompanying drawings. It is to be expressly noted that in the following description, a detailed description of known functions and designs will be omitted when it may obscure the subject matter of the present invention.
Examples
FIG. 1 is a flow chart of an embodiment of a text position detection method based on text and user representation learning according to the present invention. As shown in fig. 1, the text position detection method based on text and user representation learning of the present invention specifically comprises the following steps:
s101: preprocessing text data:
the method comprises the steps of determining a social media platform needing text position detection, collecting a text data set of a topic needing text position detection from the social media platform, wherein the text data set comprises a plurality of texts related to the topic, and an attention list and an attention-paid list among users who publish the texts.
The attention behaviors on the social media platform express one-way interests among users, the one-way interests are generated by the fact that the interests are related and the viewpoints are similar, and intuitively, the user characteristics of two people with the more similar attention lists are similar. Further, if two users are in a relationship of interest to each other, the closeness is higher than if the two users are in a one-way relationship of interest. Finally, the high probability that the same user has the same topic in a plurality of sections of texts published in a short time is the same.
Therefore, in the invention, a directed-loop-free user social relationship graph G is generated according to the attention list and the attention list among users, wherein V represents all user ID sets, E represents a set of directed edges among users, if a user i pays attention to a user j, a directed edge from the user i to the user j exists, and i, j belongs to V. And then constructing an adjacency matrix A and a degree matrix D of the user social relationship graph G ═ V, E >.
The adjacency matrix a is used for storing adjacency relations between users, and has a size of | V | × | V |, | V | representing the number of users, and an element a thereofi,jDetermined using the following formula:
Figure BDA0002959626790000051
the degree matrix D is used for storing the number of users with whom the user has contact, and is a diagonal matrix with a size of | V | × | V |, and elements on the diagonal are the number of users with whom the corresponding user has contact.
Performing Laplace matrix transformation on the adjacent matrix A and the degree matrix D to obtain a Laplace matrix LrwThe laplace matrix transformation function is as follows:
Lrw＝D-1(D-A)
where the superscript-1 represents the inversion matrix.
The position of a user in a text is normalized into three values of objection, neutrality and approval, a 3-dimensional position vector is obtained by encoding in a one-hot form, and the position vector corresponding to each text is determined as a position label vector according to the position of each text.
S102: obtaining a text vector by using a pre-trained BERT model:
unifying texts in the text data set into a preset length W: and if the text length is larger than W, deleting the excess text, and if the text length is smaller than W, filling in preset characters. Respectively inputting each text obtained by processing into a pre-trained BERT (bidirectional Encoder replication from transformations) model, and taking an output d-dimensional vector as a text vector vtWhere t represents text, t ∈ Φ, and Φ represents a set of text in the text dataset.
The BERT model is a Natural Language Processing (NLP) model proposed by Google in 2018, and is a technology which has the greatest breakthrough in the field of NLP in recent years. The model adopts an Encoder structure of a Transformer, a BERT-base model comprises 12 Encoder blocks, and a BERT-large comprises 24 Encoder blocks. The specific structure and operation of the BERT model can be found in the document "Pre-training of Deep Bidirectional transducers for Wide Understanding"
S103: constructing a vertical detection model:
in order to realize text position detection, the invention designs a position detection model. FIG. 2 is a structural diagram of a vertical inspection model according to the present invention. As shown in fig. 2, the vertical detection model in the present invention includes a GCN (Graph Convolutional neural Network) Network, an interaction layer, a decision layer, and a full connection layer, where:
the GCN is used for generating user vectors, and the specific method comprises the following steps: and taking the Laplacian matrix of the user social relationship graph as an adjacency matrix of the GCN, processing the initial vector matrix of the user by the GCN, taking the obtained output matrix as a user vector matrix, and taking each row of vectors as a user vector u corresponding to the user.
And the interaction layer is used for performing semantic fusion on the text vector and the user vector of the user to which the text obtained by the GCN belongs to obtain an interaction vector. The interaction layer comprises a hiding layer and an attention layer, wherein the hiding layer is used for performing outer product on the text vector and the user vector to obtain a d x d interaction matrix, and the interaction matrix comprises the interaction relation between the text vector and the user vector. and the attention layer is used for reducing the dimension of the interaction matrix to obtain a d-dimensional interaction vector r. Since the information dimensions of the user and the text are different, the user can be regarded as different modalities. Different modalities have different semantics, and simple splicing cannot realize effective mixing of the semantics among the modalities. Therefore, the invention sets an interaction layer, and learns the joint representation through cross-modal training, thereby realizing the feature level fusion of the user and the text.
And the decision layer is used for carrying out weighted summation on the user vector, the text vector and the interaction vector to obtain a weighted synthetic vector. The decision layer comprises a splicing layer, an MLP (multi layer Perceptron) network, a softmax layer and a weighting synthesis module, wherein the splicing layer is used for splicing the user vector, the text vector and the interaction vector to obtain a 3 x d splicing matrix. The concatenation matrix can be regarded as a length-3 sequence, and the dimension of each unit hidden layer representation on the sequence is d. MLP network for obtaining implicit expression vector u of each row vector in splicing matrixjJ is 1,2, 3. The softmax layer is used to represent the vector u according to 3 implicit representationsjDetermining a normalized weight αjThe calculation formula is as follows:
Figure BDA0002959626790000061
wherein u iswIs a d-dimensional vector, is obtained by training and is used for implicitly representing a vector ujThe dimensionality reduction is a numerical value.
A weighted synthesis module for synthesizing the weight alphajAnd carrying out weighted summation on the user vector, the text vector and the interaction vector to obtain a weighted composite vector s, wherein the weighted summation formula is as follows:
s＝∑αjhj
wherein h is1,h2,h3Representing a user vector, a text vector, and an interaction vector, respectively.
An attention mechanism is used in a decision layer, and weighted scoring is carried out on a feature level on the text feature, the user feature and the interactive feature, so as to indicate the importance difference of different features,
and the full connection layer is used for processing the weighted synthesis vector s to obtain a 3-dimensional vector, wherein each element represents the probability that the text belongs to three positions of objection, neutrality and approval.
S104: training a vertical detection model:
the laplacian matrix L in step S101 is divided into tworwAnd as an adjacent matrix of a GCN in the vertical detection model, randomly generating a d-dimensional initial vector of each user in the text data set, forming a user initial vector matrix U as a row vector, taking the user initial vector matrix U and the text vector obtained in the step S102 as input, taking a vertical label vector corresponding to the text vector as expected output, and training the vertical detection model.
In this embodiment, when performing the training of the standpoint detection model, the loss function adopts a softmax cross entropy function, the final loss is formed by adding the loss of the standpoint detection task and the regularization term, and the following objective function is jointly trained:
J＝L+λLreg
wherein L represents the cross entropy loss of the vertical detection task, LregRepresents L2Norm regularization term to prevent over fitting of neural networksAnd lambda is a hyper-parameter used for balancing the weight of the regularization term in the whole objective function.
In the training process, an Adam algorithm is adopted to optimize a loss function, and a Back Propagation (BP) algorithm is used for updating parameters. Meanwhile, in order to prevent overfitting, a Droupout training model is added before the final full-connection layer, and a nonlinear mapping function between the low-dimensional vector representation of the user and the text and the vector representation of the vertical is finally learned. For the text on the social media, the trained position detection model can be used for detecting and classifying the positions in the text.
S105: text standpoint detection:
when text published by users of a social media platform needs to be detected from the standpoint, firstly, an attention list and an attention list between current users of the social media platform are obtained, a directed acyclic user social relationship graph G ' is generated, an adjacency matrix A ' and a degree matrix D ' are constructed, and laplacian matrix L is obtained by performing laplacian matrix transformation by adopting the same method in the step S101rw′(ii) a Obtaining a d-dimensional text vector v' from the text needing to be subjected to the vertical detection by adopting the same method in the step S102;
will Laplace matrix Lrw′And as an adjacency matrix of a GCN (GCN network) in the position detection model, randomly generating d-dimensional initial vectors of all users in a user social relationship graph G ', forming a user initial vector matrix U' as row vectors, and inputting the user initial vector matrix U 'and the text vector v' into the position detection model to obtain a position detection result.
In this embodiment, a comparison experiment of the accuracy of vertical detection is performed on a Semval2016 Task6A dataset with a classical vertical detection method, wherein the dataset contains about 4,000 twitter, each twitter belongs to one topic, and there are 5 topics, namely Atheism (AT), clinical Change a relative containment (CC), Femini Movement (FM), Hillary Clinton (HC) and Legallization of infection (LA). Each tweet has a position label for one of the topics, there are three position labels, i.e., praise, objection, and neutral. The selected contrast methods include TAN (attention-based network) based text position detection method, LSTM (long short term memory network) based text position detection method, and CNN (convolutional neural network) based text position detection method. Table 1 is a comparison table of the accuracy of five topics detected by the present invention and 3 comparison methods in this embodiment.
Figure BDA0002959626790000081
TABLE 1
As shown in table 1, in five topics, the method of learning based on the user and the text representation of the present invention approaches or even exceeds the baseline of the conventional method, especially in the three topics of AT, CC and HC. The method for representing learning based on the user and the text proves that the accuracy of the position detection is improved.
Although illustrative embodiments of the present invention have been described above to facilitate the understanding of the present invention by those skilled in the art, it should be understood that the present invention is not limited to the scope of the embodiments, and various changes may be made apparent to those skilled in the art as long as they are within the spirit and scope of the present invention as defined and defined by the appended claims, and all matters of the invention which utilize the inventive concepts are protected.

Claims (1)

1. A text position detection method based on text and user representation learning is characterized by comprising the following steps:
s1: determining a social media platform needing text position detection, collecting a text data set of a topic needing text position detection from the social media platform, wherein the text data set comprises a plurality of texts related to the topic, and an attention list and an attention-to-be-paid list among users publishing the texts;
generating a directed acyclic user social relationship graph G according to the attention list and the attention list among the users<V,E>Wherein V represents the set of all user IDs, E represents the set of directed edges between users, if user i pays attention to user j, then there is a path from user i to user jToward the edge, i, j belongs to V; and then constructing a user social relationship graph G ═<V,E>An adjacency matrix a and a degree matrix D; performing Laplace matrix transformation on the adjacent matrix A and the degree matrix D to obtain a Laplace matrix Lrw；
Standardizing the position of a user in a text into three values of objection, neutrality and approval, coding in a one-hot form to obtain a 3-dimensional position vector, and determining the position vector corresponding to each text as a position label vector according to the position of each text;
s2: unifying texts in the text data set into a preset length W: if the text length is larger than W, deleting the excess text, and if the text length is smaller than W, filling preset characters; respectively inputting each text obtained by processing into a pre-trained BERT model, and taking the output d-dimensional vector as a text vector vtWherein t represents a text, t belongs to phi, phi represents a text set in the text data set, and the size of d is set according to actual needs;
s3: constructing a vertical detection model, which comprises a GCN network, an interaction layer, a decision layer and a full connection layer, wherein:
the GCN is used for generating user vectors, and the specific method comprises the following steps: taking a Laplacian matrix of a user social relationship graph as an adjacency matrix of a GCN (GCN network), processing an initial vector matrix of a user by the GCN, taking an obtained output matrix as a user vector matrix, and taking each row of vectors as user vectors corresponding to the user;
the interaction layer is used for performing semantic fusion on the text vector and a user vector of a user to which the text obtained by the GCN belongs to obtain an interaction vector; the interaction layer comprises a hidden layer and an attention layer, wherein the hidden layer is used for solving the outer product of the text vector and the user vector to obtain a dxd interaction matrix; the attention layer is used for reducing the dimension of the interaction matrix to obtain a d-dimension interaction vector;
the decision layer is used for carrying out weighted summation on the user vector, the text vector and the interaction vector to obtain a weighted synthetic vector; the decision layer comprises a splicing layer, an MLP (Multi-level hierarchical processing) network, a softmax layer and a weighting synthesis module, wherein the splicing layer is used for splicing the user vector, the text vector and the interaction vector to obtain a 3 xd splicing matrix; MLP network for acquisition stitchingImplicit representation of each row vector u in the matrixjJ is 1,2, 3; the softmax layer is used to represent the vector u according to 3 implicit representationsjDetermining a normalized weight αjThe calculation formula is as follows:
Figure FDA0002959626780000021
wherein u iswIs a d-dimensional vector for representing the implicit ujDimension reduction is a numerical value;
a weighted synthesis module for synthesizing the weight alphajAnd carrying out weighted summation on the user vector, the text vector and the interaction vector to obtain a weighted composite vector s, wherein the weighted summation formula is as follows:
s＝∑αjhj
wherein h is1,h2,h3Respectively representing a user vector, a text vector and an interaction vector;
the full-connection layer is used for processing the weighted synthesis vector s to obtain a 3-dimensional vector, wherein each element represents the probability that the text belongs to three positions of objection, neutrality and praise;
s4: the laplace matrix L in step S1 is usedrwAs an adjacency matrix of a GCN (GCN network) in the vertical detection model, randomly generating a d-dimensional initial vector of each user in the text data set and forming a user initial vector matrix U as a row vector, taking the user initial vector matrix U and the text vector obtained in the step S2 as input, taking a vertical label vector corresponding to the text vector as expected output, and training the vertical detection model;
s5: when text published by users of the social media platform needs to be detected from the standpoint, firstly, an attention list and an attention list between current users of the social media platform are obtained, a directed acyclic user social relationship graph G ' is generated, an adjacency matrix A ' and a degree matrix D ' are constructed, and laplacian matrix transformation is performed by adopting the same method of the step S1 to obtain a laplacian matrix Lrw′(ii) a Obtaining a d-dimensional text from the text to be detected from the standpoint by the same method as in step S2A vector v';
will Laplace matrix Lrw′And as an adjacency matrix of a GCN (GCN network) in the position detection model, randomly generating d-dimensional initial vectors of all users in a user social relationship graph G ', forming a user initial vector matrix U' as row vectors, and inputting the user initial vector matrix U 'and the text vector v' into the position detection model to obtain a position detection result.
CN202110233476.XA 2021-03-03 2021-03-03 Text position detection method based on text and user representation learning Active CN112949318B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110233476.XA CN112949318B (en) 2021-03-03 2021-03-03 Text position detection method based on text and user representation learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110233476.XA CN112949318B (en) 2021-03-03 2021-03-03 Text position detection method based on text and user representation learning

Publications (2)

Publication Number Publication Date
CN112949318A CN112949318A (en) 2021-06-11
CN112949318B true CN112949318B (en) 2022-03-25

Family

ID=76247341

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110233476.XA Active CN112949318B (en) 2021-03-03 2021-03-03 Text position detection method based on text and user representation learning

Country Status (1)

Country Link
CN (1) CN112949318B (en)

Citations (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108804608A (en) * 2018-05-30 2018-11-13 武汉烽火普天信息技术有限公司 A kind of microblogging rumour position detection method based on level attention
CN109299262A (en) * 2018-10-09 2019-02-01 中山大学 A kind of text implication relation recognition methods for merging more granular informations
CN109543084A (en) * 2018-11-09 2019-03-29 西安交通大学 A method of establishing the detection model of the hidden sensitive text of network-oriented social media
CN111523051A (en) * 2020-04-24 2020-08-11 山东师范大学 Social interest recommendation method and system based on graph volume matrix decomposition
CN111581401A (en) * 2020-05-06 2020-08-25 西安交通大学 Local citation recommendation system and method based on depth correlation matching
CN111966917A (en) * 2020-07-10 2020-11-20 电子科技大学 Event detection and summarization method based on pre-training language model
CA3081242A1 (en) * 2019-05-22 2020-11-22 Royal Bank Of Canada System and method for controllable machine text generation architecture
CN112001185A (en) * 2020-08-26 2020-11-27 重庆理工大学 Emotion classification method combining Chinese syntax and graph convolution neural network
CN112100388A (en) * 2020-11-18 2020-12-18 南京华苏科技有限公司 Method for analyzing emotional polarity of long text news public sentiment
WO2021000362A1 (en) * 2019-07-04 2021-01-07 浙江大学 Deep neural network model-based address information feature extraction method

Patent Citations (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108804608A (en) * 2018-05-30 2018-11-13 武汉烽火普天信息技术有限公司 A kind of microblogging rumour position detection method based on level attention
CN109299262A (en) * 2018-10-09 2019-02-01 中山大学 A kind of text implication relation recognition methods for merging more granular informations
CN109543084A (en) * 2018-11-09 2019-03-29 西安交通大学 A method of establishing the detection model of the hidden sensitive text of network-oriented social media
CA3081242A1 (en) * 2019-05-22 2020-11-22 Royal Bank Of Canada System and method for controllable machine text generation architecture
WO2021000362A1 (en) * 2019-07-04 2021-01-07 浙江大学 Deep neural network model-based address information feature extraction method
CN111523051A (en) * 2020-04-24 2020-08-11 山东师范大学 Social interest recommendation method and system based on graph volume matrix decomposition
CN111581401A (en) * 2020-05-06 2020-08-25 西安交通大学 Local citation recommendation system and method based on depth correlation matching
CN111966917A (en) * 2020-07-10 2020-11-20 电子科技大学 Event detection and summarization method based on pre-training language model
CN112001185A (en) * 2020-08-26 2020-11-27 重庆理工大学 Emotion classification method combining Chinese syntax and graph convolution neural network
CN112100388A (en) * 2020-11-18 2020-12-18 南京华苏科技有限公司 Method for analyzing emotional polarity of long text news public sentiment

Non-Patent Citations (5)

* Cited by examiner, † Cited by third party
Title
Chinese Grammatical Error Diagnosis with Graph Convolution Network and Multi-task Learning;Luo Yikang 等;《Proceedings of the 6th Workshop on Natural Language Processing Techniques for Educational Applications》;20201231;44-48 *
CNNTOP: a CNN-based Trajectory Owner Prediction Method;Luo Xucheng 等;《arXiv》;20200105;1-9 *
基于本证间隙增大的K邻域加权谱聚类算法;田文敏 等;《计算机科学与应用》;20200220;第10卷(第2期);289-302 *
基于网络社区的用户兴趣建模与推荐技术研究;李玉翔;《中国优秀博硕士学位论文全文数据库(硕士)信息科技辑》;20140215(第02期);I138-957 *
浅谈网络舆情预测方法;徐红;《科技与创新》;20191205(第23期);60-61 *

Also Published As

Publication number Publication date
CN112949318A (en) 2021-06-11

Similar Documents

Publication Publication Date Title
Zheng et al. Sentence representation method based on multi-layer semantic network
CN111061961B (en) Multi-feature-fused matrix decomposition interest point recommendation method and implementation system thereof
US9875440B1 (en) Intelligent control with hierarchical stacked neural networks
Liu et al. Deep learning with stacked denoising auto-encoder for short-term electric load forecasting
Kirilenko et al. Inter-coder agreement in one-to-many classification: fuzzy kappa
Kizielewicz et al. A new approach to identifying a multi-criteria decision model based on stochastic optimization techniques
CN112329474B (en) Attention-fused aspect-level user comment text emotion analysis method and system
Fei et al. Bidirectional grid long short-term memory (bigridlstm): A method to address context-sensitivity and vanishing gradient
Zhang et al. Forecasting hotel accommodation demand based on LSTM model incorporating internet search index
Zhou et al. Deepmove: Learning place representations through large scale movement data
Feng et al. Urban growth modeling and future scenario projection using cellular automata (CA) models and the R package optimx
Park et al. Multi-view attention network for visual dialog
Huang et al. Scenario-based marine oil spill emergency response using hybrid deep reinforcement learning and case-based reasoning
Amoudi et al. Arabic rumor detection: A comparative study
Zhang et al. Personal interest attention graph neural networks for session-based recommendation
Zhou et al. Stock prediction based on bidirectional gated recurrent unit with convolutional neural network and feature selection
Zhan et al. Variational Bayesian neural network for ensemble flood forecasting
ter Burg et al. Comparing Approaches for Explaining DNN-Based Facial Expression Classifications
Syah et al. Forecasting daily electricity price by hybrid model of fractional wavelet transform, feature selection, support vector machine and optimization algorithm
Wang et al. Short-term photovoltaic power forecasting based on a feature rise-dimensional two-layer ensemble learning model
CN113128667A (en) Cross-domain self-adaptive graph convolution balance migration learning method and system
CN112949318B (en) Text position detection method based on text and user representation learning
Medianovskyi et al. On explainable AI and abductive inference
Maldonado-Romo et al. Interoperability between real and virtual environments connected by a GAN for the path-planning problem
Wang et al. A complete reinforcement-learning-based framework for urban-safety perception

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant