CN112650861A - Personality prediction method, system and device based on task layering - Google Patents
Personality prediction method, system and device based on task layering Download PDFInfo
- Publication number
- CN112650861A CN112650861A CN202011598147.7A CN202011598147A CN112650861A CN 112650861 A CN112650861 A CN 112650861A CN 202011598147 A CN202011598147 A CN 202011598147A CN 112650861 A CN112650861 A CN 112650861A
- Authority
- CN
- China
- Prior art keywords
- vector
- personality
- user
- graph convolution
- feature
- 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.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 44
- 239000013598 vector Substances 0.000 claims abstract description 210
- 238000012549 training Methods 0.000 claims abstract description 28
- 230000004927 fusion Effects 0.000 claims abstract description 9
- 238000012545 processing Methods 0.000 claims abstract description 9
- 230000005012 migration Effects 0.000 claims abstract description 4
- 238000013508 migration Methods 0.000 claims abstract description 4
- 239000011159 matrix material Substances 0.000 claims description 25
- 238000004364 calculation method Methods 0.000 claims description 10
- 230000001149 cognitive effect Effects 0.000 claims description 8
- 230000006870 function Effects 0.000 claims description 8
- 230000019771 cognition Effects 0.000 claims description 3
- 238000007499 fusion processing Methods 0.000 claims description 3
- 230000014759 maintenance of location Effects 0.000 claims description 3
- 238000010606 normalization Methods 0.000 claims description 2
- 238000010276 construction Methods 0.000 claims 1
- 230000008569 process Effects 0.000 description 7
- 230000007246 mechanism Effects 0.000 description 4
- 238000012952 Resampling Methods 0.000 description 3
- 230000000694 effects Effects 0.000 description 3
- 238000013461 design Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000003058 natural language processing Methods 0.000 description 2
- 238000005070 sampling Methods 0.000 description 2
- 230000009466 transformation Effects 0.000 description 2
- 230000004913 activation Effects 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 238000007500 overflow downdraw method Methods 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/38—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/10—Text processing
- G06F40/12—Use of codes for handling textual entities
- G06F40/126—Character encoding
-
- 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
- G06N3/08—Learning methods
Abstract
The invention discloses a personality prediction method, a system and a device based on task layering, wherein the method comprises the following steps: performing parallel coding on user posts based on a preset BERT pre-training language model to obtain a feature vector; fusing user posts through a graph convolution network to obtain graph convolution output vectors; predicting external characteristic information of a user; completing a prediction task in a layering manner to obtain personality information; and returning the external feature vectors and the personality information to the graph convolution network and predicting again until the preset returning times are reached to obtain a personality prediction result. The system comprises: the system comprises a parallel coding module, a graph convolution network fusion module, an external data migration pre-training module, a layered self-attention personality prediction module and a message returning module. The device comprises a memory and a processor for executing the personality prediction method based on task hierarchy. By using the invention, a more accurate personality prediction result is obtained. The invention can be widely applied to the field of text processing.
Description
Technical Field
The invention relates to the field of text processing, in particular to a personality prediction method, system and device based on task layering.
Background
In deep learning research, the natural language processing direction is a subdivided field where information is encoded and mined for text content. The personality prediction problem is a cross-domain frontier problem that includes natural language processing technology and psychological knowledge, and attempts to predict the personality characteristics and personality labels of users by using texts (such as posts) published by the users on social media. In the prior art, in the process of performing early text fusion, posts of a user are spliced front and back to form a long text for feature extraction. But this introduces unnecessary or even erroneous timing information between posts, and the relevant information between posts is greatly weakened under the influence of long text processing.
Disclosure of Invention
In order to solve the above technical problems, an object of the present invention is to provide a personality prediction method, system and device based on task layering, which can independently and concurrently encode a plurality of texts, thereby improving the utilization rate of effective information and making the prediction result more accurate.
The first technical scheme adopted by the invention is as follows: a personality prediction method based on task layering comprises the following steps:
acquiring user posts and carrying out parallel coding on the user posts based on a preset BERT pre-training language model to obtain a feature vector;
fusing user posts through a graph convolution network, constructing a topological graph based on the characteristic vector and obtaining a graph convolution output vector;
predicting external feature information of a user according to the topological graph based on a pre-trained external feature model to obtain an external feature vector;
completing a personality prediction task of four dimensions in a layering mode according to the external feature vector and the graph convolution output vector to obtain personality information;
and returning the external characteristic information and the personality information to the graph convolution network and predicting again until the preset returning times are reached to obtain a personality prediction result.
Further, the step of obtaining the user posts and performing parallel coding on the user posts based on a preset BERT pre-training language model to obtain the feature vectors specifically includes:
setting the sample processing number, the maximum value of the number of posts belonging to the user and the vector of the number of effective posts by taking the user as a sample unit to obtain preset information;
constructing a two-dimensional vector corresponding to a user according to preset information;
and parallelly transmitting the two-dimensional vectors corresponding to the user into a preset BERT pre-training language model for coding to obtain the feature vectors of the posts.
Further, the step of fusing the user posts through a graph convolution network, constructing a topological graph based on the feature vectors, and obtaining a graph convolution output vector specifically includes:
and taking the user posts as graph nodes, constructing a corresponding topological graph for each user sample based on the feature vector to which the user posts belong, and obtaining a convolved output vector.
Further, let feature vector EmbedijThe Pmax X d dimension node feature matrix is X, the adjacent matrix between the nodes is A, and the calculation method of the convolved output vector is as follows:
wherein
I is a unit matrix of the image data,
is that
Degree matrix of (H)(0)＝X，W(l)Is a parameter matrix, the convolution layer number l is 2, the adjacent matrix A is a random initialization symmetric matrix,
the operation is on the adjacency matrixPerforming a Laplace normalization transform, said H(l+1)And representing the output vector after the node feature convolution.
Further, the method for constructing the pre-trained external feature model specifically comprises the following steps:
acquiring an external data set containing user post data and tags of user age, gender and the like, and training an external feature model according to the user post data, the corresponding tags of the user age and the gender to obtain model parameters;
and fixing the model parameters to obtain a first layer of external feature model with the function of predicting the age and the gender of the user.
Further, the step of completing the personality prediction task of four dimensions in a layering manner according to the external feature vector and the graph convolution output vector to obtain personality information specifically comprises:
based on a second-layer encoder, performing fusion processing on the graph convolution output vector, the age vector and the gender vector to complete an attention direction prediction task and a judgment mode prediction task to obtain a second-layer feature vector;
outputting a vector, an age vector, a gender vector and a second-layer feature vector to a graph convolution based on a third-layer encoder, and completing a life style prediction task and a cognitive style prediction task to obtain a third-layer feature vector;
and obtaining a personality information prediction result of the user according to the second layer feature vector and the third layer feature vector.
Further, the second-layer feature vector comprises an attention direction feature vector and a judgment mode feature vector, and the third-layer feature vector comprises a life mode feature vector and a cognitive mode feature vector.
Further, the step of returning the external feature vector and the personality information to the graph convolution network and predicting again until reaching the preset returning times to obtain the personality prediction result specifically includes:
respectively calculating the return retention weights of the age vector, the gender vector, the attention direction characteristic vector, the judgment mode characteristic vector, the life mode characteristic vector and the cognition mode characteristic vector;
calculating according to the return reservation weight of the feature vector to obtain a reservation message vector;
the reserved message vector and the graph convolution vector are transmitted back to the graph convolution network and the graph convolution network is updated, and the updated graph convolution network is obtained;
and predicting the personality information again based on the updated graph convolution network until the preset number of return times is reached to obtain a personality prediction result.
The second technical scheme adopted by the invention is as follows: a personality prediction system based on task hierarchy, comprising:
the parallel coding module is used for acquiring user posts and carrying out parallel coding on the user posts based on a preset BERT pre-training language model to obtain feature vectors;
the graph convolution network fusion module is used for fusing the user posts through a graph convolution network, constructing a topological graph based on the characteristic vector and obtaining a graph convolution output vector;
the external data migration pre-training module is used for predicting external feature information of the user according to the topological graph based on a pre-trained external feature model to obtain an external feature vector;
the hierarchical self-attention personality prediction module is used for completing the personality prediction tasks of four dimensions in a hierarchical mode according to the external feature vectors and the graph convolution output vectors to obtain personality information;
and the message returning module is used for returning the external characteristic vector and the personality information to the graph convolution network and predicting again until the preset returning times are reached to obtain a personality prediction result.
The third technical scheme adopted by the invention is as follows: a personality prediction device based on task hierarchy, comprising:
at least one processor;
at least one memory for storing at least one program;
when executed by the at least one processor, cause the at least one processor to implement a personality prediction method based on task hierarchy as described above.
The method, the system and the device have the advantages that: the invention can independently and parallelly encode a plurality of texts in the text encoding stage, can combine the feature vectors of different posts, excavate the correlation between posts about user character features and personality information, abstract deeper text semantics, improve the utilization rate of effective information, and divide the tasks with different prediction difficulties into different levels by dividing the prediction task of each dimension of the personality into levels, so that the task with better and simpler prediction result can complete training more quickly, and the result is transmitted to the next layer to provide additional reliable information for the task prediction of the next layer, thereby obtaining more accurate prediction result.
Drawings
FIG. 1 is a flowchart illustrating the steps of task hierarchy based personality prediction in accordance with the present invention;
FIG. 2 is a block diagram of a personality prediction system based on task hierarchy according to the present invention;
FIG. 3 is a schematic diagram of a model architecture of an embodiment of the present invention;
FIG. 4 is a flow chart of data processing according to an embodiment of the present invention.
Detailed Description
The invention is described in further detail below with reference to the figures and the specific embodiments. The step numbers in the following embodiments are provided only for convenience of illustration, the order between the steps is not limited at all, and the execution order of each step in the embodiments can be adapted according to the understanding of those skilled in the art.
Referring to fig. 1 and 3, the present invention provides a personality prediction method based on task hierarchy, which includes the following steps:
s1, obtaining user posts and carrying out parallel coding on the user posts based on a preset BERT pre-training language model to obtain feature vectors;
s2, fusing the user posts through a graph convolution network, constructing a topological graph based on the feature vectors and obtaining graph convolution output vectors;
s3, predicting external feature information of the user according to the topological graph based on the pre-trained external feature model to obtain an external feature vector;
s4, completing four-dimensional personality prediction tasks in a layering mode according to the external feature vectors and the graph convolution output vectors to obtain personality information;
and S5, returning the external characteristic information and the personality information to the graph convolution network and predicting again until the preset returning times are reached to obtain a personality prediction result.
As a preferred embodiment of the method, the step of obtaining the user posts and performing parallel coding on the user posts based on a preset BERT pre-training language model to obtain the feature vectors specifically includes:
setting the sample processing number, the maximum value of the number of posts belonging to the user and the vector of the number of effective posts by taking the user as a sample unit to obtain preset information;
constructing a two-dimensional vector corresponding to a user according to preset information;
and parallelly transmitting the two-dimensional vectors corresponding to the user into a preset BERT pre-training language model for coding to obtain the feature vectors of the posts.
Specifically, a dimension reduction method is adopted, the number of samples processed by the model each time is set to be batch _ size, the maximum value of the number of posts to which the user belongs is Pmax, a vector post _ mask is set, and the number of effective posts of a single sample is recorded. If the number of single sample posts does not reach Pmax, then use "[ pad]The "value is filled in. Let the jth post of the ith sample be postij。
Will be [ batch _ size, Pmax ] originally]Is reduced to [ batch _ size ] Pmax]The vectors are parallelly transmitted into a large-scale pre-training language model BERT for coding to obtain a d-dimensional feature vector (Embed) of each postij)：
Embedij＝BERT(postij)，d＝768
Each feature vector can reflect semantic information of a corresponding post in the calculation process of the high-dimensional space. Wherein post filled is referred to according to post _ maskijIs the eigenvector EmbeddedijThe weight is set to 0 in the subsequent operation. After the coding operation is finished, the size is [ batch _ size ] Pmax [ ]]The vector of (b) is expanded to [ batch _ size, Pmax [ ]]The two-dimensional vector of (2) thus satisfies the requirement of encoding single posts in parallel by taking the user as a unit.
Further, as a preferred embodiment of the method, the step of fusing the user posts through a graph convolution network, constructing a topological graph based on the feature vectors, and obtaining a graph convolution output vector specifically includes:
and taking the user posts as graph nodes, constructing a corresponding topological graph for each user sample based on the feature vector to which the user posts belong, and obtaining a convolved output vector.
Further, as a preferred embodiment of the method, an eigenvector Embed is setijThe Pmax X d dimension node feature matrix is X, the adjacent matrix between the nodes is A, and the calculation method of the convolved output vector is as follows:
wherein
I is a unit matrix of the image data,
is that
Degree matrix of (H)(0)＝X，W(l)Is a parameter matrix, the convolution layer number l is 2, the adjacent matrix A is a random initialization symmetric matrix,
the operation is to perform a laplace normalized transform on the adjacency matrix. After the convolution calculation is finished, the obtained characteristic information in the adjacent matrix A represents the correlation information between the nodes. Node characteristic matrix volumeThe integrated output vector H(l+1)Then, each node fuses the feature vectors of the related nodes adjacent to itself. Specifically, the personality prediction task is to try to find the topological relation among nodes by taking a user as an object and each post as a node through graph convolution calculation so as to integrally depict the personality information of the user.
Further, as a preferred embodiment of the method, the method for constructing the pre-trained external feature model specifically includes the following steps:
acquiring an external data set containing user post data and tags of user age, gender and the like, and training an external feature model according to the user post data, the corresponding tags of the user age and the gender to obtain model parameters;
and fixing the model parameters to obtain a first layer of external feature model with the function of predicting the age and the gender of the user.
Specifically, as the first layer of the hierarchical model, the model is migrated to the task of personality prediction, and by predicting age and gender first and then combining these characteristic information, a more accurate prediction about personality characteristics is obtained.
In addition, in the later personality prediction process, in order to ensure that the age and gender prediction capability is not influenced by the personality prediction data set, a resampling mechanism is introduced, sampling is carried out on the external data set again, training of age and gender prediction is carried out, the prediction capability of the model is ensured to be maintained at a high accuracy rate all the time, the use effect of the external features is ensured, and then training of personality prediction is carried out on the model continuously.
As a preferred embodiment of the method, the step of completing the personality prediction task of four dimensions hierarchically according to the external feature vector and the graph convolution output vector to obtain the personality information specifically includes:
based on a second-layer encoder, performing fusion processing on the graph convolution output vector, the age vector and the gender vector to complete an attention direction prediction task and a judgment mode prediction task to obtain a second-layer feature vector;
outputting a vector, an age vector, a gender vector and a second-layer feature vector to a graph convolution based on a third-layer encoder, and completing a life style prediction task and a cognitive style prediction task to obtain a third-layer feature vector;
and obtaining a personality information prediction result of the user according to the second layer feature vector and the third layer feature vector.
Specifically, the second and third layers of the model are based on the four-dimensional personality prediction task of the MBTI index.
The two prediction tasks corresponding to the second layer are an IE task (attention direction) and a TF task (judgment mode), and by utilizing corresponding encoders IE _ Encoder and TF _ Encoder, the image convolution output vector and the age vector and the gender vector of the previous layer are taken as input, and the information of the lower layer is fused, so that the feature vectors of the two tasks of the current layer are obtained. The respective classifiers (IE _ Classifier, TF _ Classifier) are further trained to obtain predicted IE and TF prediction results.
The two predictive tasks corresponding to the third layer are the JP task (lifestyle) and the NS task (cognitive style). The principle of this layer is substantially the same as that of the second layer.
In the process of fusing vectors output by different characteristics and modules, the module adopts a self-attention mechanism to perform coding calculation. And performing parallel splicing on the graph convolution output vector and the low-level feature vectors (age, gender, IE, TF and the like) to obtain a vector X, and performing linear transformation to obtain a query vector Q, a key vector K and a value vector V.
Q＝WQH0
K＝WKH0
V＝WVH0
And obtaining a final code vector through scaling and nonlinear transformation:
wherein d iskIs the eigenvector dimension used to perform the scaling of the dot product.
In the module, through a self-attention mechanism, the model dynamically learns the weights of different connections among vectors, and can more flexibly process the relation of feature vectors of different modules and different levels. And the design of the hierarchical model separates tasks with different difficulties into different levels for calculation, so that the data flow is more efficient, and the model has better interpretability.
Further as a preferred embodiment of the method, the second-layer feature vector includes an attention direction feature vector and a judgment mode feature vector, and the third-layer feature vector includes a life mode feature vector and a cognitive mode feature vector.
Further, as a preferred embodiment of the method, the step of returning the external feature vector and the personality information to the graph convolution network and predicting again until the preset number of return times is reached to obtain the personality prediction result specifically includes:
in particular, based on 6 feature vectors of different prediction tasks, respectively an age vector hageGender vector hgenderIE eigenvectors hIETF feature vector hTFJP feature vector hJPAnd NS feature vector hNS. The module designs a message return mechanism, and the feature vector obtained in one turn and the graph convolution vector h are multipliedGCNAnd under the condition of mutual dominance, the interaction and the update of the data are carried out again.
Respectively calculating the return retention weights of the age vector, the gender vector, the attention direction characteristic vector, the judgment mode characteristic vector, the life mode characteristic vector and the cognition mode characteristic vector;
calculating the return reserve weight of each feature vector:
where o belongs to { age, genre, IE, TF, JP, NS }.
Calculating according to the return reservation weight of the feature vector to obtain a reservation message vector;
the final reservation message vector is:
the reserved message vector and the graph convolution vector are transmitted back to the graph convolution network and the graph convolution network is updated, and the updated graph convolution network is obtained;
next, the graph is convolved with a vector hGCNCombined message vector hoAnd carrying out the next updating:
h′GCN＝fmp(hGCN：ho)
wherein [:]indicating a splicing operation, fmpThe calculation includes a fully connected network layer and a RELU activation function.
Updated vector hGCNAnd then the feature vectors are transmitted into the hierarchical model module again for training to obtain a new round of feature vectors of each prediction task. The message returning operation can control error transmission, information sharing and fusion can be performed among different tasks, information originally on a high layer of the model can be utilized by a low layer, and different positive feedback effects are obtained under the control of a loss function. In the actual operation process of the model, a fixed number of pass-back times can be set.
And predicting the personality information again based on the updated graph convolution network until the preset number of return times is reached to obtain a personality prediction result.
The specific data processing and model training and prediction processes in this patent are described below:
a series of hyper-parameters, such as an age/gender task accuracy threshold P and a message return time N, are predefined. And D, defining the current processing data set as D, wherein D can be used for taking the original personality prediction data set and the external age and gender data set. The flow chart refers to fig. 4:
1) model initialization, data set D is external age and gender data set, initial message return times n is 0, identifier Flag is False
2) The data is sampled from D and processed to a length and format acceptable to the model.
3) And the graph convolution text fusion module obtains a text fusion vector by adopting a fusion method of parallel coding of a BERT pre-training language model and graph convolution.
4) And training self-attention encoders of the age task and the gender task according to the fusion vector, and comparing the data labels to obtain the accuracy rate p of the two tasks.
5) It is determined whether the identifier Flag is true. If true, the feature vectors are passed into the next layer model. Otherwise, judging whether the accuracy rate P reaches a preset threshold value P.
6) If P does not reach the preset threshold P, the gradient updates the parameters of the first tier age and gender encoder. If the threshold has been met or exceeded, indicating that the first layer has been trained sufficiently, the data set D is replaced with a personality predicted data set and the identifier is set to true.
7) When the identifier Flag is true, the feature vector is transmitted to a higher-level model, and IE/TF prediction (attention direction/judgment mode) and JP/NS prediction (life mode/cognitive mode) are performed once to obtain feature vectors of respective dimensions.
8) And judging whether the current return times N are less than N. And if the sum of the weights is less than N, adding 1 to N, carrying out message return calculation, carrying out weight analysis on the feature vector, and carrying out iteration by combining the graph convolution vector. If N is larger than or equal to N, the message is returned and the personality prediction parameters of the second layer and the third layer of the model are updated according to the current prediction result.
9) And judging whether to perform resampling operation. And if resampling is carried out, modifying the sampling data set into an external age and gender data set, and retraining the prediction capability of the age and gender of the model and consolidating the parameters.
10) The objective loss function used is focal loss. And continuously updating the model parameters according to the steps until the prediction result is converged.
As shown in fig. 2, a personality prediction system based on task hierarchy includes:
the parallel coding module is used for acquiring user posts and carrying out parallel coding on the user posts based on a preset BERT pre-training language model to obtain feature vectors;
the graph convolution network fusion module is used for fusing the user posts through a graph convolution network, constructing a topological graph based on the characteristic vector and obtaining a graph convolution output vector;
the external data migration pre-training module is used for predicting external feature information of the user according to the topological graph based on a pre-trained external feature model to obtain an external feature vector;
the hierarchical self-attention personality prediction module is used for completing the personality prediction tasks of four dimensions in a hierarchical mode according to the external feature vectors and the graph convolution output vectors to obtain personality information;
and the message returning module is used for returning the external characteristic information and the personality information to the graph convolution network and predicting again until the preset returning times are reached to obtain a personality prediction result.
The contents in the above method embodiments are all applicable to the present system embodiment, the functions specifically implemented by the present system embodiment are the same as those in the above method embodiment, and the beneficial effects achieved by the present system embodiment are also the same as those achieved by the above method embodiment.
A personality prediction device based on task hierarchy:
at least one processor;
at least one memory for storing at least one program;
when executed by the at least one processor, cause the at least one processor to implement a personality prediction method based on task hierarchy as described above.
The contents in the above method embodiments are all applicable to the present apparatus embodiment, the functions specifically implemented by the present apparatus embodiment are the same as those in the above method embodiments, and the advantageous effects achieved by the present apparatus embodiment are also the same as those achieved by the above method embodiments.
While the preferred embodiments of the present invention have been illustrated and described, it will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention as defined by the appended claims.
Claims (10)
1. A personality prediction method based on task layering is characterized by comprising the following steps:
acquiring user posts and carrying out parallel coding on the user posts based on a preset BERT pre-training language model to obtain a feature vector;
fusing user posts through a graph convolution network, constructing a topological graph based on the characteristic vector and obtaining a graph convolution output vector;
predicting external feature information of a user according to the topological graph based on a pre-trained external feature model to obtain an external feature vector;
completing a personality prediction task of four dimensions in a layering mode according to the external feature vector and the graph convolution output vector to obtain personality information;
and returning the external feature vectors and the personality information to the graph convolution network and predicting again until the preset returning times are reached to obtain a personality prediction result.
2. The task hierarchy-based personality prediction method of claim 1, wherein the step of obtaining user posts and encoding the user posts in parallel based on a pre-defined BERT pre-training language model to obtain feature vectors comprises:
setting the sample processing number, the maximum value of the number of posts belonging to the user and the vector of the number of effective posts by taking the user as a sample unit to obtain preset information;
constructing a two-dimensional vector corresponding to a user according to preset information;
and parallelly transmitting the two-dimensional vectors corresponding to the user into a preset BERT pre-training language model for coding to obtain the feature vectors of the posts.
3. The task hierarchy-based personality prediction method of claim 2, wherein the step of fusing the user posts through a graph convolution network, constructing a topological graph based on the feature vectors and obtaining a graph convolution output vector comprises:
and taking the user posts as graph nodes, constructing a corresponding topological graph for each user sample based on the feature vector to which the user posts belong, and obtaining a convolved output vector.
4. The personality prediction method based on task hierarchy of claim 3, wherein an eigenvector Embed is setijThe Pmax X d dimension node feature matrix is X, the adjacent matrix between the nodes is A, and the calculation method of the convolved output vector is as follows:
wherein
I is a unit matrix of the image data,
is that
Degree matrix of (H)(0)＝X，W(l)Is a parameter matrix, the convolution layer number l is 2, the adjacent matrix A is a random initialization symmetric matrix,
the operation is to perform a Laplace normalization transform on the adjacency matrix, H(l+1)And representing the output vector after the node feature convolution.
5. The personality prediction method based on task layering as claimed in claim 4, wherein the construction method of the pre-trained external feature model specifically comprises the following steps:
acquiring an external data set containing user post data and tags of user age, gender and the like, and training an external feature model according to the user post data, the corresponding tags of the user age and the gender to obtain model parameters;
and fixing the model parameters to obtain a first layer of external feature model with the function of predicting the age and the gender of the user.
6. The personality prediction method based on task layering according to claim 5, wherein the step of completing the personality prediction task of four dimensions hierarchically according to the external feature vector and the graph convolution output vector to obtain personality information specifically comprises:
based on a second-layer encoder, performing fusion processing on the graph convolution output vector, the age vector and the gender vector to complete an attention direction prediction task and a judgment mode prediction task to obtain a second-layer feature vector;
outputting a vector, an age vector, a gender vector and a second-layer feature vector to a graph convolution based on a third-layer encoder, and completing a life style prediction task and a cognitive style prediction task to obtain a third-layer feature vector;
and obtaining a personality information prediction result of the user according to the second layer feature vector and the third layer feature vector.
7. The personality prediction method based on task layering of claim 6, wherein the second-layer feature vectors comprise attention direction feature vectors and judgment mode feature vectors, and the third-layer feature vectors comprise life mode feature vectors and cognitive mode feature vectors.
8. The personality prediction method based on task hierarchy of claim 7, wherein the step of returning the extrinsic feature vectors and the personality information to the graph convolution network and re-predicting until reaching a preset number of return times to obtain the personality prediction result specifically comprises:
respectively calculating the return retention weights of the age vector, the gender vector, the attention direction characteristic vector, the judgment mode characteristic vector, the life mode characteristic vector and the cognition mode characteristic vector;
calculating according to the return reservation weight of the feature vector to obtain a reservation message vector;
the reserved message vector and the graph convolution vector are transmitted back to the graph convolution network and the graph convolution network is updated, and the updated graph convolution network is obtained;
and predicting the personality information again based on the updated graph convolution network until the preset number of return times is reached to obtain a personality prediction result.
9. A personality prediction system based on task hierarchy, comprising:
the parallel coding module is used for acquiring user posts and carrying out parallel coding on the user posts based on a preset BERT pre-training language model to obtain feature vectors;
the graph convolution network fusion module is used for fusing the user posts through a graph convolution network, constructing a topological graph based on the characteristic vector and obtaining a graph convolution output vector;
the external data migration pre-training module is used for predicting external feature information of the user according to the topological graph based on a pre-trained external feature model to obtain an external feature vector;
the hierarchical self-attention personality prediction module is used for completing the personality prediction tasks of four dimensions in a hierarchical mode according to the external feature vectors and the graph convolution output vectors to obtain personality information;
and the message returning module is used for returning the external characteristic vector and the personality information to the graph convolution network and predicting again until the preset returning times are reached to obtain a personality prediction result.
10. A personality prediction device based on task hierarchy, comprising:
at least one processor;
at least one memory for storing at least one program;
when executed by the at least one processor, cause the at least one processor to implement a task hierarchy-based personality prediction method as claimed in any one of claims 1-7.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011598147.7A CN112650861A (en) | 2020-12-29 | 2020-12-29 | Personality prediction method, system and device based on task layering |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011598147.7A CN112650861A (en) | 2020-12-29 | 2020-12-29 | Personality prediction method, system and device based on task layering |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN112650861A true CN112650861A (en) | 2021-04-13 |
Family
ID=75363925
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011598147.7A Pending CN112650861A (en) | 2020-12-29 | 2020-12-29 | Personality prediction method, system and device based on task layering |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112650861A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113837445A (en) * | 2021-08-27 | 2021-12-24 | 合肥工业大学 | Personality prediction method and system based on attention mechanism |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109829154A (en) * | 2019-01-16 | 2019-05-31 | 中南民族大学 | Semantic-based personality prediction technique, user equipment, storage medium and device |
| CN110473592A (en) * | 2019-07-31 | 2019-11-19 | 广东工业大学 | The multi-angle of view mankind for having supervision based on figure convolutional network cooperate with lethal gene prediction technique |
| KR20200040652A (en) * | 2018-10-10 | 2020-04-20 | 고려대학교 산학협력단 | Natural language processing system and method for word representations in natural language processing |
| CN111259976A (en) * | 2020-01-21 | 2020-06-09 | 中山大学 | Personality detection method based on multi-mode alignment and multi-vector representation |
| CN111897908A (en) * | 2020-05-12 | 2020-11-06 | 中国科学院计算技术研究所 | Event extraction method and system fusing dependency information and pre-training language model |
| CN111914885A (en) * | 2020-06-19 | 2020-11-10 | 合肥工业大学 | Multitask personality prediction method and system based on deep learning |
-
2020
- 2020-12-29 CN CN202011598147.7A patent/CN112650861A/en active Pending
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20200040652A (en) * | 2018-10-10 | 2020-04-20 | 고려대학교 산학협력단 | Natural language processing system and method for word representations in natural language processing |
| CN109829154A (en) * | 2019-01-16 | 2019-05-31 | 中南民族大学 | Semantic-based personality prediction technique, user equipment, storage medium and device |
| CN110473592A (en) * | 2019-07-31 | 2019-11-19 | 广东工业大学 | The multi-angle of view mankind for having supervision based on figure convolutional network cooperate with lethal gene prediction technique |
| CN111259976A (en) * | 2020-01-21 | 2020-06-09 | 中山大学 | Personality detection method based on multi-mode alignment and multi-vector representation |
| CN111897908A (en) * | 2020-05-12 | 2020-11-06 | 中国科学院计算技术研究所 | Event extraction method and system fusing dependency information and pre-training language model |
| CN111914885A (en) * | 2020-06-19 | 2020-11-10 | 合肥工业大学 | Multitask personality prediction method and system based on deep learning |
Non-Patent Citations (1)
| Title |
|---|
| MOHAMMAD HOSSEIN AMIRHOSSEINI 等: ""Machine Learning Approach to Personality Type Prediction Based on the Myers–Briggs Type Indicator"", 《MULTIMODAL TECHNOL. INTERACT.》, vol. 4, no. 1 * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113837445A (en) * | 2021-08-27 | 2021-12-24 | 合肥工业大学 | Personality prediction method and system based on attention mechanism |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111522962B (en) | Sequence recommendation method, device and computer readable storage medium | |
| CN111737476B (en) | Text processing method and device, computer readable storage medium and electronic equipment | |
| CN111950269A (en) | Text statement processing method and device, computer equipment and storage medium | |
| CN116415654A (en) | Data processing method and related equipment | |
| CN113905391A (en) | Ensemble learning network traffic prediction method, system, device, terminal, and medium | |
| CN111914085A (en) | Text fine-grained emotion classification method, system, device and storage medium | |
| CN108665506A (en) | Image processing method, device, computer storage media and server | |
| CN111858898A (en) | Text processing method and device based on artificial intelligence and electronic equipment | |
| WO2023231513A1 (en) | Conversation content generation method and apparatus, and storage medium and terminal | |
| CN116521882A (en) | Domain length text classification method and system based on knowledge graph | |
| Cao et al. | Stacked residual recurrent neural network with word weight for text classification | |
| CN115687610A (en) | Text intention classification model training method, recognition device, electronic equipment and storage medium | |
| CN115114409A (en) | Civil aviation unsafe event combined extraction method based on soft parameter sharing | |
| US11941360B2 (en) | Acronym definition network | |
| CN114444476A (en) | Information processing method, apparatus and computer readable storage medium | |
| CN112650861A (en) | Personality prediction method, system and device based on task layering | |
| CN111259673B (en) | Legal decision prediction method and system based on feedback sequence multitask learning | |
| CN116737897A (en) | Intelligent building knowledge extraction model and method based on multiple modes | |
| CN116384379A (en) | Chinese clinical term standardization method based on deep learning | |
| CN112132269B (en) | Model processing method, device, equipment and storage medium | |
| CN114444517A (en) | Intelligent law judgment method for numerical perception with increased sentencing standard knowledge | |
| CN110414012B (en) | Artificial intelligence-based encoder construction method and related equipment | |
| CN114282058A (en) | Method, device and equipment for model training and video theme prediction | |
| CN113627192A (en) | Relation extraction method and device based on two-layer convolutional neural network | |
| CN114626378A (en) | Named entity recognition method and device, electronic equipment and computer readable storage medium |
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 |