CN115033700A - Cross-domain emotion analysis method, device and equipment based on mutual learning network - Google Patents
Cross-domain emotion analysis method, device and equipment based on mutual learning network Download PDFInfo
- Publication number
- CN115033700A CN115033700A CN202210954299.9A CN202210954299A CN115033700A CN 115033700 A CN115033700 A CN 115033700A CN 202210954299 A CN202210954299 A CN 202210954299A CN 115033700 A CN115033700 A CN 115033700A
- Authority
- CN
- China
- Prior art keywords
- text data
- mutual learning
- target
- feature
- representations corresponding
- 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
- 230000008451 emotion Effects 0.000 title claims abstract description 143
- 238000004458 analytical method Methods 0.000 title claims abstract description 112
- 230000006870 function Effects 0.000 claims abstract description 103
- 239000013598 vector Substances 0.000 claims abstract description 98
- 238000012549 training Methods 0.000 claims abstract description 63
- 238000000034 method Methods 0.000 claims abstract description 36
- 238000005457 optimization Methods 0.000 claims abstract description 15
- 238000013508 migration Methods 0.000 claims description 135
- 230000005012 migration Effects 0.000 claims description 135
- 238000000605 extraction Methods 0.000 claims description 78
- 238000004422 calculation algorithm Methods 0.000 claims description 36
- 150000001875 compounds Chemical class 0.000 claims description 27
- 230000002457 bidirectional effect Effects 0.000 claims description 26
- 238000012545 processing Methods 0.000 claims description 21
- 230000002996 emotional effect Effects 0.000 claims description 11
- 238000001514 detection method Methods 0.000 claims description 8
- 230000004913 activation Effects 0.000 claims description 6
- 238000010606 normalization Methods 0.000 claims description 6
- 238000007405 data analysis Methods 0.000 claims description 4
- 239000011159 matrix material Substances 0.000 claims description 3
- 230000004044 response Effects 0.000 abstract description 3
- 230000008569 process Effects 0.000 description 15
- 238000004590 computer program Methods 0.000 description 8
- 238000010586 diagram Methods 0.000 description 6
- 238000004891 communication Methods 0.000 description 3
- 230000008878 coupling Effects 0.000 description 3
- 238000010168 coupling process Methods 0.000 description 3
- 238000005859 coupling reaction Methods 0.000 description 3
- 230000008034 disappearance Effects 0.000 description 3
- 125000004122 cyclic group Chemical group 0.000 description 2
- 238000004880 explosion Methods 0.000 description 2
- 241000288105 Grus Species 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000013500 data storage Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000003058 natural language processing Methods 0.000 description 1
- 238000009877 rendering Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 239000000126 substance Substances 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
- G06F16/35—Clustering; Classification
- G06F16/353—Clustering; Classification into predefined classes
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/205—Parsing
- G06F40/211—Syntactic parsing, e.g. based on context-free grammar [CFG] or unification grammars
-
- 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
- G06N3/08—Learning methods
Abstract
The invention relates to the field of emotion analysis, in particular to a cross-domain emotion analysis method, a cross-domain emotion analysis device, a cross-domain emotion analysis equipment and a storage medium based on a mutual learning network, wherein a training text data set is obtained, the training text data set is input into a preset word embedding model, a word embedding vector set is obtained, the word embedding vector set is input into the preset mutual learning network, a loss function corresponding to the mutual learning network is constructed, optimization training is carried out, a target mutual learning network is obtained, in response to an analysis instruction, text data to be analyzed are obtained, the text data to be analyzed are input into the target mutual learning network, and emotion analysis results output by the target mutual learning network are obtained. Compared with the prior art, the method and the device have the advantages that the emotion analysis is carried out on the text data set to be detected more comprehensively, and the accuracy and the efficiency of the emotion analysis are improved.
Description
Technical Field
The invention relates to the field of emotion analysis, in particular to a cross-domain emotion analysis method, a cross-domain emotion analysis device, cross-domain emotion analysis equipment and a storage medium based on a mutual learning network.
Background
Cross-domain emotion classification is one of important tasks in natural language processing, the purpose of the task is to learn knowledge from Source domain (Source domain) data with abundant labels to guide Target domain (Target domain) training with few labels or no labels, and the task has important value on field research with labels lacking. Most of the existing correlation models relate to a feature extractor and an emotion classifier, wherein the feature extractor is responsible for learning the invariant features from two domains, and the emotion classifier is trained on the source domain only to guide the learning of the feature extractor. Often such models ignore the target domain emotional polarity and do not exploit the target domain emotional polarity. In addition, the existing cross-domain emotion classification method neglects the characteristics of complex syntactic structure and various semantic information of texts in different domains.
In order to enable a feature extractor to have the capability of learning a domain invariant representation, the features of a source domain and a target domain need to be aligned, a more mainstream method at present is to use a countermeasure network as a domain discriminator, and use a gradient inversion layer to deceive the feature extractor so that the feature extractor cannot distinguish which domain data comes from, however, the countermeasure network does not have a definite loss to indicate a training process in an actual training process, and only can depend on experience to judge whether a generator and a discriminator reach dynamic balance, which leads to unstable training of the countermeasure network and easily causes the problem of gradient disappearance.
Disclosure of Invention
Based on the method, the device, the equipment and the storage medium, the mutual learning network is constructed by utilizing a deep mutual learning method, each group of mutual learning migration channels in the mutual learning network are associated with the source field and the target field, the difference between the two fields is minimized, the mutual learning network is optimized and trained, the accuracy and the efficiency of optimization training are improved, and the emotion analysis can be carried out on text data more comprehensively.
The technical method comprises the following steps:
in a first aspect, an embodiment of the present application provides a cross-domain emotion analysis method based on a mutual learning network, including the following steps:
acquiring a training text data set, wherein the training text data set comprises a text data set of a source field and a text data set of a target field, the text data set comprises a plurality of text data, the text data comprises a plurality of sentences, and the sentences comprise a plurality of words;
inputting the training text data set into a preset word embedding model to obtain a word embedding vector set, wherein the word embedding vector set comprises word embedding vector representations corresponding to text data of a plurality of source fields and word embedding vector representations corresponding to text data of a plurality of target fields;
inputting the word embedding vector set into a preset mutual learning network, constructing a loss function corresponding to the mutual learning network, performing optimization training, and obtaining a target mutual learning network, wherein the mutual learning network comprises two groups of mutual learning migration channels;
and responding to an analysis instruction, acquiring text data to be analyzed, inputting the text data to be analyzed into the target mutual learning network, and acquiring an emotion analysis result output by the target mutual learning network.
In a second aspect, an embodiment of the present application provides a cross-domain emotion analysis apparatus based on a mutual learning network, including:
the training text data set acquisition module is used for acquiring a training text data set, wherein the training text data set comprises text data of a plurality of source fields and text data of a plurality of target fields, the text data comprises a plurality of sentences, and the sentences comprise a plurality of words;
a word embedding vector representation obtaining module, configured to input the training text data set into a preset word embedding model, and obtain a word embedding vector set, where the word embedding vector set includes word embedding vector representations corresponding to text data of a plurality of source fields and word embedding vector representations corresponding to text data of a plurality of target fields;
the network training module is used for inputting the word embedding vector set into a preset mutual learning network, constructing a loss function corresponding to the mutual learning network, performing optimization training and obtaining a target mutual learning network, wherein the mutual learning network comprises two groups of mutual learning migration channels;
and the text data analysis module is used for responding to an analysis instruction, acquiring text data to be analyzed, inputting the text data to be analyzed into the target mutual learning network, and acquiring an emotion analysis result output by the target mutual learning network.
In a third aspect, an embodiment of the present application provides a computer device, including: a processor, a memory, and a computer program stored on the memory and executable on the processor; the computer program when executed by the processor implements the steps of the cross-domain emotion analysis method based on a mutual learning network according to the first aspect.
In a fourth aspect, the present application provides a storage medium storing a computer program, which when executed by a processor implements the steps of the cross-domain emotion analysis method based on mutual learning network according to the first aspect.
In this embodiment, a cross-domain emotion analysis method, a cross-domain emotion analysis device, a cross-domain emotion analysis equipment and a storage medium based on a mutual learning network are provided, a mutual learning network is constructed by using a deep mutual learning method, each group of mutual learning migration channels in the mutual learning network are associated with a source domain and a target domain, and the mutual learning network is optimally trained by minimizing the difference between the two domains, so that the accuracy and efficiency of the optimal training are improved, and emotion analysis can be performed on text data more comprehensively.
For a better understanding and practice, the invention is described in detail below with reference to the accompanying drawings.
Drawings
FIG. 1 is a schematic flowchart of a cross-domain emotion analysis method based on a mutual learning network according to an embodiment of the present application;
FIG. 2 is a schematic flowchart of S3 in the cross-domain emotion analysis method based on mutual learning network according to an embodiment of the present application;
fig. 3 is a schematic flowchart of S31 in the cross-domain emotion analysis method based on mutual learning network according to an embodiment of the present application;
fig. 4 is a schematic flowchart of S311 in the cross-domain emotion analysis method based on the mutual learning network according to an embodiment of the present application;
fig. 5 is a schematic flowchart of S312 in the cross-domain emotion analysis method based on mutual learning network according to an embodiment of the present application;
FIG. 6 is a schematic flowchart of S32 in the cross-domain emotion analysis method based on mutual learning network according to an embodiment of the present application;
FIG. 7 is a schematic flowchart of S33 in the cross-domain emotion analysis method based on mutual learning network according to an embodiment of the present application;
fig. 8 is a schematic flowchart of S34 in the cross-domain emotion analysis method based on mutual learning network according to an embodiment of the present application;
FIG. 9 is a schematic flowchart of S4 in the cross-domain emotion analysis method based on mutual learning network according to an embodiment of the present application;
FIG. 10 is a schematic structural diagram of a cross-domain emotion analysis device based on a mutual learning network according to an embodiment of the present application;
fig. 11 is a schematic structural diagram of a computer device according to an embodiment of the present application.
Detailed Description
Reference will now be made in detail to the exemplary embodiments, examples of which are illustrated in the accompanying drawings. When the following description refers to the accompanying drawings, like numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present application. Rather, they are merely examples of apparatus and methods consistent with certain aspects of the present application, as detailed in the appended claims.
The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. As used in this application and the appended claims, the singular forms "a", "an", and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It should also be understood that the term "and/or" as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items.
It is to be understood that although the terms first, second, third, etc. may be used herein to describe various information, such information should not be limited to these terms. These terms are only used to distinguish one type of information from another. For example, first information may also be referred to as second information, and similarly, second information may also be referred to as first information, without departing from the scope of the present application. The word "if/if" as used herein may be interpreted as "at … …" or "when … …" or "in response to a determination", depending on the context.
Referring to fig. 1, fig. 1 is a schematic flowchart of a cross-domain emotion analysis method based on a mutual learning network according to an embodiment of the present application, including the following steps:
s1: a training text data set is obtained.
The main execution body of the cross-domain emotion analysis method based on the mutual learning network is analysis equipment (hereinafter referred to as analysis equipment) of the cross-domain emotion analysis method based on the mutual learning network.
In an alternative embodiment, the analysis device may be a computer device, a server, or a server cluster formed by combining a plurality of computer devices.
In this embodiment, the analysis device may obtain a training text data set input by a user, and also obtain a corresponding training text data set from a preset database, where the training text data set includes a text data set of the source field and a text data set of the target field, the text data set includes a plurality of text data, the text data includes a plurality of sentences, and the sentences include a plurality of words.
S2: and inputting the training text data set into a preset word embedding model to obtain a word embedding vector set.
The Word embedding model is a Word2vec Word embedding model and is used for obtaining Word embedding vectors corresponding to each text data.
In this embodiment, the analysis device inputs the training text data set into a preset Word embedding model, maps each Word in the text data into a low-dimensional vector space, and obtains a Word embedding vector set by querying a pre-trained Word2Vec matrix, where the Word embedding vector set includes Word embedding vector representations corresponding to text data of a plurality of source fields and Word embedding vector representations corresponding to text data of a plurality of target fields.
S3: and inputting the word embedding vector set into a preset mutual learning network, constructing a loss function corresponding to the mutual learning network, performing optimization training, and acquiring a target mutual learning network.
The mutual learning network comprises two groups of mutual learning migration channels, and the two groups of mutual learning migration channels comprise a feature extraction module, an emotion classification module, a field difference learning module and a label detection module.
In this embodiment, the analysis device inputs the word embedded vector set to a preset mutual learning network, constructs a loss function corresponding to the mutual learning network, performs optimization training, and obtains a target mutual learning network.
Referring to fig. 2, fig. 2 is a schematic flowchart of S3 in the cross-domain emotion analysis method based on mutual learning network according to an embodiment of the present application, including steps S31 to S35, which are as follows:
s31: and respectively inputting the word embedding vector set to the feature extraction modules in the two groups of mutual learning migration channels, and acquiring text feature representations corresponding to the text data of the plurality of source fields and the text feature representations corresponding to the text data of the plurality of target fields, which are output by the feature extraction modules of the two groups of mutual learning migration channels.
In this embodiment, the analysis device inputs the word embedding vector set to the feature extraction modules in the two groups of mutual learning migration channels, and obtains text feature representations corresponding to the text data of the plurality of source fields and text feature representations corresponding to the text data of the plurality of target fields, which are output by the feature extraction modules in the two groups of mutual learning migration channels.
The feature extraction module includes a semantic feature extraction module and a syntactic feature extraction module which are connected in sequence, please refer to fig. 3, and fig. 3 is a schematic flow diagram of S31 in the cross-domain emotion analysis method based on the mutual learning network according to an embodiment of the present application, including steps S311 to S313, which are as follows:
s311: and respectively inputting the word embedding vector set to a semantic feature extraction module in the corresponding feature extraction module, and respectively acquiring semantic feature representations corresponding to the text data of the plurality of source fields and the text data of the plurality of target fields, which are output by the semantic feature extraction modules of the two groups of mutual learning migration channels.
In this embodiment, the analysis device inputs the word embedding vector set to the semantic feature extraction modules in the corresponding feature extraction modules, and obtains semantic feature representations corresponding to the text data of the plurality of source domains and the text data of the plurality of target domains output by the semantic feature extraction modules of the two groups of mutual learning migration channels, respectively.
The semantic feature extraction module includes a first bidirectional gated loop unit, a soft attention unit, a second bidirectional gated loop unit, and a convolution attention unit, where the convolution attention unit includes a plurality of convolution layers, please refer to fig. 4, and fig. 4 is a schematic flow diagram of S311 in the cross-domain emotion analysis method based on the mutual learning network according to an embodiment of the present application, including steps S3111 to S3114, which are as follows:
s3111: and inputting the word embedding vector set into a first bidirectional gating circulating unit in the corresponding semantic feature extraction module, respectively coding the word embedding vector set, and acquiring hidden layer feature representations corresponding to the text data of the plurality of source fields and the hidden layer feature representations corresponding to the text data of the plurality of target fields, which are output by the bidirectional gating circulating units of the semantic feature extraction modules of the two groups of mutual learning migration channels.
In this embodiment, the analysis device inputs the word embedding vector set into the corresponding first bidirectional gating cyclic unit in the semantic feature extraction module for encoding processing, and obtains two sub-hidden layer feature representations output by two GRUs in different directions, which are specifically as follows:
in the formula (I), the compound is shown in the specification,
、
respectively sub-hidden layer feature representations in two different directions,nfor the number of words in the sentence,iis the first in a sentence in the text dataiThe number of the individual words is,jas the first in the text datajThe number of the sentences is one,
is the coding function of the bi-directional gated cyclic unit,
embedding a vector representation for the word;
and combining the sub-hidden layer feature representations respectively to obtain hidden layer feature representations corresponding to the text data of the plurality of source fields and the hidden layer feature representations corresponding to the text data of the plurality of target fields, which are output by a bidirectional gating circulation unit of a semantic feature extraction module of the two groups of mutual learning migration channels, wherein the hidden layer feature representations are as follows:
in the formula (I), the compound is shown in the specification,
is the hidden layer feature representation.
S3112: inputting hidden layer feature representations corresponding to the text data of the source fields and hidden layer feature representations corresponding to the text data of the target fields into corresponding soft attention units in the semantic feature extraction module, acquiring corresponding attention weight parameters and attention weight parameters corresponding to the text data of the target fields according to a preset attention weight parameter calculation algorithm, and respectively acquiring sentence feature representations corresponding to the text data of the source fields and sentence feature representations corresponding to the text data of the target fields, which are output by the soft attention units of the two groups of mutual learning migration channels according to a preset sentence feature representation calculation algorithm.
In order to measure the contribution of each word in each sentence in the text data to the emotional information of the sentence, in this embodiment, the analyzing device inputs the hidden layer feature representations corresponding to the text data of the plurality of source fields and the hidden layer feature representations corresponding to the text data of the plurality of target fields to the corresponding Soft Attention unit (Soft-Attention) in the semantic feature extraction module, and obtains the corresponding Attention weight parameter and the Attention weight parameter corresponding to the text data of the plurality of target fields according to a preset Attention weight parameter calculation algorithm, where the Attention weight parameter calculation algorithm is:
in the formula (I), the compound is shown in the specification,
for the purpose of the attention weight parameter,
for a preset first trainable network parameter,
for a preset second trainable network parameter,
is a preset third trainable network parameter;
and according to a preset sentence characteristic representation calculation algorithm, respectively acquiring sentence characteristic representations corresponding to the text data of the source fields output by the soft attention units of the two groups of mutual learning migration channels and sentence characteristic representations corresponding to the text data of the target fields, so as to improve the accuracy of the acquired sentence characteristic representations and further improve the effect of optimization training of the mutual learning network, wherein the sentence characteristic representation calculation algorithm is as follows:
in the formula (I), the compound is shown in the specification,
is a sentence feature representation.
S3113: and respectively inputting sentence characteristic representations corresponding to the text data of the source fields and sentence characteristic representations corresponding to the text data of the target fields into corresponding second bidirectional gating circulation units in the semantic characteristic extraction module for coding, and respectively acquiring the sentence characteristic representations corresponding to the text data of the source fields and the sentence characteristic representations corresponding to the text data of the target fields after coding and processing, which are output by the second bidirectional gating circulation units of the two groups of mutual learning migration channels.
In this embodiment, the analysis device inputs the sentence feature representations corresponding to the text data of the plurality of source domains and the sentence feature representations corresponding to the text data of the plurality of target domains into the corresponding second bidirectional gating circulation units in the semantic feature extraction module respectively for encoding, and obtains the sentence feature representations corresponding to the text data of the plurality of source domains and the sentence feature representations corresponding to the text data of the plurality of target domains after encoding, which are output by the second bidirectional gating circulation units of the two groups of mutual learning migration channels, respectively, wherein the sentence feature representations after encoding are:
in the formula (I), the compound is shown in the specification,
for the sentence feature representation after the encoding process, BiGRU () is an encoding function of the second bidirectional gating loop unit.
S3114: and inputting sentence characteristic representations corresponding to the text data of the plurality of source fields after the coding processing and sentence characteristic representations corresponding to the text data of the plurality of target fields after the coding processing into corresponding convolution attention units in the semantic characteristic extraction module, respectively performing weighting processing on the sentence characteristic representations corresponding to the text data of the plurality of source fields after the coding processing and the sentence characteristic representations corresponding to the text data of the plurality of target fields after the coding processing according to a preset attention algorithm, and acquiring semantic characteristic representations corresponding to the text data of the plurality of source fields and semantic characteristic representations corresponding to the text data of the plurality of target fields.
The convolution Attention unit (CNN-Attention) is used for giving different weights to each input sentence feature representation, extracting more key and important information and enabling the network to make more accurate judgment.
In this embodiment, the analysis device inputs the sentence feature representations corresponding to the text data of the plurality of source fields after the encoding process and the sentence feature representations corresponding to the text data of the plurality of target fields after the encoding process to the convolution attention unit in the semantic feature extraction module, respectively performs weighting processing on the sentence feature representations corresponding to the text data of the plurality of source fields after the encoding process and the sentence feature representations corresponding to the text data of the plurality of target fields after the encoding process according to a preset attention algorithm, and acquires semantic feature representations corresponding to the text data of the plurality of source fields
And semantic feature representations corresponding to the text data of the target domains
Taking the semantic feature representation corresponding to the text data of the acquired target domain as an example, the operation process of the attention algorithm is as follows:
in the formula (I), the compound is shown in the specification,qfor the number of windows in the convolutional layer,mfor the number of sentences in the text data,tis the first in the convolutional layertA convolution filter is set in the filter to filter the convolution,
in order to represent the characteristics of the convolution,
in order to be a non-linear activation function,
is a firsttThe convolution function of each of the convolution filters,
which represents a convolution operation, the operation of the convolution,
for the purpose of the preset convolution coefficients,
is a firstkSemantic feature representations corresponding to the text data of the individual target domains.
S312: and respectively inputting the word embedding vector set to a syntactic feature extraction module in the corresponding feature extraction modules, and respectively acquiring syntactic feature representations corresponding to the text data of the plurality of source fields and the text data of the plurality of target fields, which are output by the syntactic feature extraction modules of the two groups of mutual learning migration channels.
In this embodiment, the analysis device inputs the word embedding vector set to the syntactic feature extraction modules in the corresponding feature extraction modules, and obtains syntactic feature representations corresponding to the text data of the plurality of source fields and the text data of the plurality of target fields, which are output by the syntactic feature extraction modules of the two groups of mutual learning migration channels, respectively.
The syntactic feature extraction module comprises a bidirectional gating circulation unit, a multilayer graph attention network unit and a convolution unit, wherein the multilayer graph attention network unit comprises a plurality of graph attention network layers; referring to fig. 5, fig. 5 is a schematic flowchart of S312 in the cross-domain emotion analysis method based on the mutual learning network according to an embodiment of the present application, including steps S3121 to S3123, which are specifically as follows:
s3121: and inputting the word embedding vector set to a corresponding bidirectional gating circulating unit in the syntactic feature extraction module, and respectively acquiring the context hidden layer feature representations corresponding to the text data of the plurality of source fields and the context hidden layer feature representations corresponding to the text data of the plurality of target fields, which are output by the bidirectional gating circulating units of the syntactic feature extraction modules of the two groups of mutual learning migration channels.
The syntactic characteristic extracting module mainly considers that syntactic dependencies also exist among words of different sentences, so that the analysis equipment treats the whole text data as a long sentence, and each text data containsndA word, wherein,nd=m*n。
in this embodiment, the analyzing device inputs the word embedding vector set to the corresponding bidirectional gating cycle unit in the syntactic feature extraction module, and obtains the context hidden layer feature representations corresponding to the text data in the plurality of source fields and the context hidden layer feature representations corresponding to the text data in the plurality of target fields, which are output by the bidirectional gating cycle units of the syntactic feature extraction modules of the two groups of mutual learning migration channels, respectively, as follows:
in the formula (I), the compound is shown in the specification,
for the context-hiding layer feature representation,BiGRU() And coding functions of bidirectional gating cycle units in the syntactic feature extraction module.
S3122: and inputting the context hidden layer feature representation corresponding to the text data of the source fields and the context hidden layer feature representation corresponding to the text data of the target fields into the corresponding multilayer drawing attention network unit in the syntactic feature extraction module, taking the context hidden layer feature representation as a state vector of a first drawing attention network layer of the multilayer drawing attention network unit, and respectively acquiring state vectors corresponding to nodes of the plurality of drawing attention network layers of the two groups of mutual learning migration channels according to a preset state vector calculation algorithm.
The multi-layer graph ATTENTION network unit is GAT (GRAPH ATTENTION NETWORKS), and is a multi-layer graph ATTENTION network constructed in a stacking mode.
In this embodiment, the analysis device inputs the context hidden layer feature representations corresponding to the text data of the source fields and the context hidden layer feature representations corresponding to the text data of the target fields into corresponding multilayer graph attention network units in the syntax feature extraction module, uses the context hidden layer feature representations as state vectors of a first graph attention network layer of the multilayer graph attention network units, and respectively obtains state vectors corresponding to nodes of the graph attention network layers of the two groups of mutual learning migration channels according to a preset state vector calculation algorithm, where the state vector calculation algorithm is:
in the formula (I), the compound is shown in the specification,
is a non-linear activation function;
is at the firstlPersonal attention network layer, the firstxA node and ayRaw attention score data between individual nodes,
for a pre-set trainable weight vector, | | | is expressed as a concatenation operation,
for a pre-set trainable weight matrix,
is as followslFirst of a layerxThe state vector corresponding to each node is calculated,
after normalization processing, in the second steplPersonal attention network layer, the firstxA node and ayAn attention weight parameter between the individual nodes,
in order to be a non-linear activation function,
for all and nodesxA set of adjacent nodes.
S3123: and inputting the state vectors corresponding to the nodes of the plurality of graph attention network layers of the two groups of mutual learning migration channels into the corresponding convolution units in the syntactic feature extraction module for feature extraction, and acquiring syntactic feature representations corresponding to the text data of the plurality of source fields and syntactic feature representations corresponding to the text data of the plurality of target fields.
In this embodiment, the analysis device inputs the state vectors corresponding to the nodes of the plurality of graph attention network layers of the two groups of channels that are learned and migrated mutually into the corresponding convolution units in the syntactic feature extraction module to perform feature extraction, and obtains syntactic feature representations corresponding to the text data of the plurality of source fields and syntactic feature representations corresponding to the text data of the plurality of target fields.
S313: and respectively splicing the semantic feature representation and the syntactic feature representation corresponding to the text data of the source fields and the semantic feature representation and the syntactic feature representation corresponding to the text data of the target fields, and acquiring the text feature representation corresponding to the text data of the source fields and the text feature representation corresponding to the text data of the target fields, which are output by the syntactic feature extraction modules of the two groups of mutual learning migration channels.
In this embodiment, the analysis device respectively represents the semantic features, the syntactic features and the destinations corresponding to the text data of the source fieldsAnd splicing semantic feature representation and syntactic feature representation corresponding to the text data of the mark domain, and acquiring text feature representations corresponding to the text data of the source fields output by the syntactic feature extraction modules of the two groups of mutual learning migration channels
And text characteristic representation corresponding to the text data of the target fields
。
S32: and respectively inputting the text feature representations corresponding to the text data of the source fields and the text feature representations corresponding to the text data of the target fields into emotion classification modules of the two groups of mutual learning migration channels, obtaining emotion feature representations corresponding to the text data of the source fields of the two groups of mutual learning migration channels and emotion feature representations corresponding to the text data of the target fields, and constructing a first loss function of the two groups of mutual learning migration channels according to the emotion feature representations corresponding to the text data of the source fields.
In this embodiment, the analysis device inputs the text feature representations corresponding to the text data of the source fields and the text feature representations corresponding to the text data of the target fields to the emotion classification modules of the two groups of mutual learning migration channels, obtains emotion feature representations corresponding to the text data of the source fields of the two groups of mutual learning migration channels and emotion feature representations corresponding to the text data of the target fields, and constructs a first loss function of the two groups of mutual learning migration channels according to the emotion feature representations corresponding to the text data of the source fields.
Referring to fig. 6, fig. 6 is a schematic flowchart of S32 in the cross-domain emotion analysis method based on the mutual learning network according to an embodiment of the present application, including steps S321 to S322, which are as follows:
s321: and obtaining emotion feature representations corresponding to the text data of the source fields of the two groups of mutual learning migration channels and emotion feature representations corresponding to the text data of the target fields respectively according to the text feature representations corresponding to the text data of the source fields, the text feature representations corresponding to the text data of the target fields and an emotion classification module of the two groups of mutual learning migration channels by using a preset emotion feature representation calculation algorithm.
The emotional feature representation calculation algorithm comprises the following steps:
in the formula (I), the compound is shown in the specification,
representing corresponding emotional features of the text data of the source fields,
the emotional characteristic representation corresponding to the text data of the target fields,
in order to be a function of the normalization,
for a preset fourth trainable network parameter,
is a preset fifth trainable network parameter;
in this embodiment, the analysis device obtains, according to the text feature representations corresponding to the text data of the plurality of source fields, the text feature representations corresponding to the text data of the plurality of target fields, and the emotion classification module of the two groups of mutual learning migration channels, the emotion feature representation calculation algorithm preset in the emotion classification module, the emotion feature representation calculation algorithm corresponding to the text data of the plurality of source fields of the two groups of mutual learning migration channels, and the emotion feature representation corresponding to the text data of the plurality of target fields.
S322: and obtaining label data corresponding to the text data of the plurality of source fields, and constructing a first loss function of the two groups of mutual learning migration channels according to the emotional feature representation and the label data corresponding to the text data of the plurality of source fields.
For the emotional feature representations corresponding to the text data of the source fields, in this embodiment, the analysis device obtains tag data corresponding to the text data of the source fields, and constructs a first loss function of the two groups of mutual learning migration channels by calculating cross entropy according to the emotional feature representations and the tag data corresponding to the text data of the source fields, where the first loss function is:
in the formula (I), the compound is shown in the specification,
for the purpose of said first loss function,
the number of text data of the source domain is set to the text data of the source domain,
is as followskLabel data corresponding to the text data of the source field.
S33: respectively inputting the text feature representations corresponding to the text data of the target domains into the label detection modules of the two groups of mutual learning migration channels, acquiring the label feature representations corresponding to the text data of the target domains of the two groups of mutual learning migration channels, inputting the label feature representations corresponding to the text data of the target domains of the mutual learning migration channels into the emotion classification module of the other group of mutual learning migration channels, and constructing a second loss function of the two groups of mutual learning migration channels according to the emotion feature representations corresponding to the text data of the source domains and the label feature representations corresponding to the text data of the target domains.
In this embodiment, the analysis device respectively inputs text feature representations corresponding to the text data of the target domains into the tag detection modules of the two groups of mutual learning migration channels, obtains tag feature representations corresponding to the text data of the target domains of the two groups of mutual learning migration channels, inputs tag feature representations corresponding to the text data of the target domains of the group of mutual learning migration channels into the emotion classification module of the other group of mutual learning migration channels, and constructs the second loss function of the two groups of mutual learning migration channels according to the emotion feature representations corresponding to the text data of the source domains and the tag feature representations corresponding to the text data of the target domains.
Referring to fig. 7, fig. 7 is a schematic flowchart of S33 in the cross-domain emotion analysis method based on the mutual learning network according to an embodiment of the present application, including steps S331 to S332, which are as follows:
s331: and acquiring label feature representations corresponding to the text data of the plurality of target domains of the two groups of mutual learning migration channels according to the text feature representations corresponding to the text data of the plurality of target domains and a preset label feature representation calculation algorithm in the label detection modules of the two groups of mutual learning migration channels.
The label feature representation calculation algorithm is as follows:
in the formula (I), the compound is shown in the specification,
label feature representations corresponding to the text data of the several target fields,
for the preset sixth trainable network parameters,
is a preset seventh trainable network parameter;
s332: and inputting label feature representations corresponding to the text data of a plurality of target domains of one group of mutual learning migration channels into an emotion classification module of another group of mutual learning migration channels, and constructing a second loss function of the two groups of mutual learning migration channels according to the emotion feature representations corresponding to the text data of the plurality of source domains and the label feature representations corresponding to the text data of the plurality of target domains.
The second loss function is:
in the formula (I), the compound is shown in the specification,
、
respectively learning second loss functions corresponding to the migration channels for the two groups
、
The emotion feature representations corresponding to the text data of the target domains corresponding to the two groups of mutual learning migration channels are respectively displayed,
、
label features corresponding to the text data of the target domain corresponding to the two groups of mutual learning migration channels respectivelyThe representation of the sign is that,
is a preset divergence coefficient.
In this embodiment, the analysis device inputs label feature representations corresponding to text data of a plurality of target domains of one group of mutual learning migration channels into an emotion classification module of another group of mutual learning migration channels, and constructs a second loss function of the two groups of mutual learning migration channels according to emotion feature representations corresponding to text data of a plurality of source domains and label feature representations corresponding to text data of a plurality of target domains.
S34: and inputting the text feature representations corresponding to the text data of the source fields and the text feature representations corresponding to the text data of the target fields into a field difference learning module of the quantity group mutual learning migration channels, and constructing a third loss function of the two groups of mutual learning migration channels.
In this embodiment, the analysis device inputs the text feature representations corresponding to the text data of the plurality of source domains and the text feature representations corresponding to the text data of the plurality of target domains into the domain difference learning module of the two groups of mutual learning migration channels, and constructs a third loss function of the two groups of mutual learning migration channels.
Referring to fig. 8, fig. 8 is a schematic flowchart of S34 in the cross-domain emotion analysis method based on the mutual learning network according to an embodiment of the present application, which includes steps S341 to S343 as follows:
s341: and combining the text characteristic representations corresponding to the text data of the source fields and the text characteristic representations corresponding to the text data of the target fields respectively to obtain the text characteristic representation corresponding to the text data set of the source fields and the text characteristic representation corresponding to the text data set of the target fields.
In this embodiment, the analysis device combines the text feature representations corresponding to the text data of the plurality of source fields and the text feature representations corresponding to the text data of the plurality of target fields, respectively, to obtain the text feature representations corresponding to the text data sets of the source fields, which are used as the representation distribution of the source fields; and the text characteristic representation corresponding to the text data set of the target domain is used as the representation distribution of the target domain.
S342: and inputting the text characteristic representation corresponding to the text data set of the source field and the text characteristic representation corresponding to the text data set of the target field into a field difference learning module of the quantity group mutual learning migration channel to construct a bulldozer distance function.
The dozer distance function is:
in the formula (I), the compound is shown in the specification,
as a function of the bulldozer distance,
the number of text data of the source domain is set to the text data of the source domain,
is the number of text data of the target domain in the text data set of the target domain,
the source domain is represented by a field of view,
a target domain is represented by a target field,
for the learning function in the domain difference learning module,
is the text of the source domainThe corresponding text feature representation of the present data set,
and representing the text characteristic corresponding to the text data set of the target domain.
In this embodiment, the analysis device inputs the text feature representation corresponding to the text data set of the source domain and the text feature representation corresponding to the text data set of the target domain into the domain difference learning module of the quantitative group mutual learning migration channel, and constructs a bulldozer distance function. The difference between the representation distribution of the source domain and the representation distribution of the target domain is estimated by adopting the calculated bulldozer distance (Wasserstein), so that the difference between the two domains is minimized, the domain invariant features are extracted, and the accuracy of network training is improved.
S343: and constructing a gradient penalty function according to preset learning training parameters and balance coefficients, and acquiring a third loss function of the two field difference learning modules of the mutual learning migration channels according to the bulldozer distance function and the gradient penalty function.
In order to improve the accuracy of network training, after each gradient update, the analysis device may execute Lipschitz constraint by clipping the weight within a preset clipping range, where the weight clipping may cause problems of capacity insufficiency and gradient disappearance or explosion, and therefore, in this embodiment, the analysis device constructs a gradient penalty function according to preset learning training parameters and balance coefficients to avoid problems of capacity insufficiency and gradient disappearance or explosion, subtracts the bulldozer distance function from the gradient penalty function, and obtains a function subtraction result as a third loss function of the two field difference learning modules of the mutual learning migration channels.
Wherein the gradient penalty function is:
in the formula (I), the compound is shown in the specification,
is a gradient penalty characteristic representation
And with
The serial connection of random points on the feature space;
is the equilibrium coefficient;
the third loss function is:
s35: and respectively constructing total loss functions corresponding to the two groups of mutual learning migration channels according to the first loss function, the first loss function and the third loss function of the two groups of mutual learning migration channels, and performing optimization training according to the total loss functions corresponding to the two groups of mutual learning migration channels to obtain the target mutual learning network.
In this embodiment, the analysis device respectively constructs a total loss function corresponding to the two groups of mutual learning migration channels according to a first loss function, and a third loss function of the two groups of mutual learning migration channels, where the total loss function is:
in the formula (I), the compound is shown in the specification,
、
respectively learning total loss functions corresponding to the migration channels for the two groups,
、
respectively learning first loss functions corresponding to the migration channels for the two groups,
、
respectively learning second loss functions corresponding to the migration channels for the two groups,
、
respectively learning third loss functions corresponding to the migration channels for the two groups,
is a first hyper-parameter that is pre-set,
is a preset second hyper-parameter.
Carrying out optimization training according to the total loss functions corresponding to the two groups of mutual learning migration channels to obtain the target mutual learning network, specifically, solving by the analysis equipment
Wherein, in the step (A),
learning the training parameters for a preset second function to estimate the dozer distance, the analysis device first optimizing by iterative learning feature representation since the dozer distance is continuously differentiable almost anywhere, and after the optimization is completed, the analysis deviceOver-fixing the second function learning training parameters and minimizing the bulldozer distance to extract common features between the fields by solving the minimum maximum problem, namely:
wherein the content of the first and second substances,
the training parameters are learned for a preset second function. In a preferred embodiment, the analysis device should set the balance coefficient to 0 when performing the minimization operation, since the gradient penalizes the training process that should not be guided.
S4: and responding to an analysis instruction, acquiring text data to be analyzed, inputting the text data to be analyzed into the target mutual learning network, and acquiring an emotion analysis result output by the target mutual learning network.
The analysis instruction is sent by a user and received by the analysis equipment.
In this embodiment, the analysis device responds to an analysis instruction, obtains text data to be analyzed, inputs the text data to be analyzed into the target mutual learning network, and obtains an emotion analysis result output by the target mutual learning network.
Referring to fig. 9, fig. 9 is a schematic flowchart of S4 in the cross-domain emotion analysis method based on mutual learning network according to an embodiment of the present application, including steps S41 to S42, as follows:
s41: inputting the text data to be analyzed into the target mutual learning network to obtain text feature representation of the text data to be analyzed, and obtaining the emotion polarity vector of the text data to be analyzed according to the text feature representation of the text data to be analyzed and a preset emotion polarity vector calculation algorithm.
The calculation algorithm of the emotion polarity vector comprises the following steps:
in the formula (I), the compound is shown in the specification,pfor the said vector of the polarity of emotion,
in order to be a function of the normalization,
in order to update the parameters for the weights,
the bias term is updated for the weight(s),
representing the text characteristics of the text data to be analyzed;
in this embodiment, the analysis device inputs the text data to be analyzed into the target mutual learning network, obtains text feature representation of the text data to be analyzed, and obtains an emotion polarity vector of the analyzed text data according to the text feature representation of the analyzed text data and a preset emotion polarity vector calculation algorithm.
S42: and acquiring the emotion polarity corresponding to the dimensionality with the maximum probability according to the emotion classification polarity probability distribution vector, and using the emotion polarity as an emotion analysis result output by the target mutual learning network.
In this embodiment, the analysis device obtains, according to the emotion classification polarity probability distribution vector, the emotion polarity corresponding to the dimension with the highest probability as the emotion analysis result output by the target mutual learning network, specifically, when the emotion polarity is obtained through calculation
=[
1，
2，
3，
4，
5]=[0.1，0.5，0.1，0.2，0.1]Maximum probability
And 2, the emotion polarity score corresponding to the dimension with the highest probability is 2, and as the emotion analysis result output by the target mutual learning network, the higher the score is, the more positive the text is, and the lower the score is, the more negative the text is.
Referring to fig. 10, fig. 10 is a schematic structural diagram of a cross-domain emotion analyzing apparatus based on mutual learning network according to an embodiment of the present application, where the apparatus may implement all or a part of the cross-domain emotion analyzing method based on mutual learning network through software, hardware or a combination of the two, and the apparatus 7 includes:
a training text data set obtaining module 101, configured to obtain a training text data set, where the training text data set includes text data of a plurality of source fields and text data of a plurality of target fields, the text data includes a plurality of sentences, and the sentences include a plurality of words;
a word embedding vector representation obtaining module 102, configured to input the training text data set into a preset word embedding model, and obtain a word embedding vector set, where the word embedding vector set includes word embedding vector representations corresponding to text data of a plurality of source fields and word embedding vector representations corresponding to text data of a plurality of target fields;
the network training module 103 is configured to input the word embedding vector set into a preset mutual learning network, construct a loss function corresponding to the mutual learning network, perform optimization training, and obtain a target mutual learning network, where the mutual learning network includes two groups of mutual learning migration channels;
the text data analysis module 104 is configured to, in response to an analysis instruction, acquire text data to be analyzed, input the text data to be analyzed into the target mutual learning network, and acquire an emotion analysis result output by the target mutual learning network.
In an embodiment of the application, a training text data set is obtained through a training text data set obtaining module, wherein the training text data set comprises text data of a plurality of source fields and text data of a plurality of target fields, the text data comprises a plurality of sentences, and the sentences comprise a plurality of words; the word embedding vector representation obtaining module is used for inputting the training text data set into a preset word embedding model to obtain a word embedding vector set, wherein the word embedding vector set comprises word embedding vector representations corresponding to text data of a plurality of source fields and word embedding vector representations corresponding to text data of a plurality of target fields; inputting the word embedding vector set into a preset mutual learning network through a network training module, constructing a loss function corresponding to the mutual learning network, performing optimization training, and obtaining a target mutual learning network, wherein the mutual learning network comprises two groups of mutual learning migration channels; the text data analysis module responds to an analysis instruction to obtain text data to be analyzed, the text data to be analyzed is input into the target mutual learning network, and emotion analysis results output by the target mutual learning network are obtained. By utilizing a deep mutual learning method, a mutual learning network is constructed, each group of mutual learning migration channels in the mutual learning network is associated with a source field and a target field, and the mutual learning network is optimally trained by minimizing the difference between the two fields, so that the accuracy and efficiency of the optimal training are improved, and the emotion analysis can be more comprehensively carried out on the text data.
Referring to fig. 11, fig. 11 is a schematic structural diagram of a computer device according to an embodiment of the present application, where the computer device 11 includes: a processor 111, a memory 112, and a computer program 113 stored on the memory 112 and operable on the processor 111; the computer device may store a plurality of instructions, where the instructions are suitable for being loaded by the processor 111 and executing the method steps in the embodiments described in fig. 1 to 9, and a specific execution process may refer to specific descriptions of the embodiments described in fig. 1 to 9, which is not described herein again.
The Memory 112 may include a Random Access Memory (RAM) 112, and may also include a Read-Only Memory (Read-Only Memory) 112. Optionally, the memory 112 includes a non-transitory computer-readable medium. The memory 112 may be used to store instructions, programs, code sets, or instruction sets. The memory 112 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch instructions, etc.), instructions for implementing the various method embodiments described above, and the like; the storage data area may store data and the like referred to in the above respective method embodiments. The memory 112 may optionally be at least one storage device located remotely from the processor 111.
The embodiment of the present application further provides a storage medium, where the storage medium may store a plurality of instructions, where the instructions are suitable for being loaded by a processor and being executed in the method steps of the first to third embodiments, and a specific execution process may refer to specific descriptions of the embodiments shown in fig. 1 to fig. 9, which is not described herein again.
It will be apparent to those skilled in the art that, for convenience and brevity of description, only the above-mentioned division of the functional units and modules is illustrated, and in practical applications, the above-mentioned function distribution may be performed by different functional units and modules according to needs, that is, the internal structure of the apparatus is divided into different functional units or modules to perform all or part of the above-mentioned functions. Each functional unit and module in the embodiments may be integrated in one processing unit, or each unit may exist alone physically, or two or more units are integrated in one unit, and the integrated unit may be implemented in a form of hardware, or in a form of software functional unit. In addition, specific names of the functional units and modules are only for convenience of distinguishing from each other, and are not used for limiting the protection scope of the present application. For the specific working processes of the units and modules in the system, reference may be made to the corresponding processes in the foregoing method embodiments, which are not described herein again.
In the above embodiments, the description of each embodiment has its own emphasis, and reference may be made to the related description of other embodiments for parts that are not described or recited in any embodiment.
Those of ordinary skill in the art will appreciate that the various illustrative elements and algorithm steps described in connection with the embodiments disclosed herein may be implemented as electronic hardware or combinations of computer software and electronic hardware. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the implementation. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present invention.
In the embodiments provided in the present invention, it should be understood that the disclosed apparatus/terminal device and method may be implemented in other ways. For example, the above-described embodiments of the apparatus/terminal device are merely illustrative, and for example, the division of the modules or units is only one logical division, and there may be other divisions when actually implemented, for example, a plurality of units or components may be combined or integrated into another system, or some features may be omitted, or not executed. In addition, the shown or discussed mutual coupling or direct coupling or communication connection may be an indirect coupling or communication connection through some interfaces, devices or units, and may be in an electrical, mechanical or other form.
The units described as separate parts may or may not be physically separate, and parts displayed as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of the embodiment.
In addition, functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may exist alone physically, or two or more units are integrated into one unit. The integrated unit may be implemented in the form of hardware, or may also be implemented in the form of a software functional unit.
The integrated modules/units, if implemented in the form of software functional units and sold or used as separate products, may be stored in a computer readable storage medium. Based on such understanding, all or part of the flow of the method according to the embodiments of the present invention may also be implemented by a computer program, which may be stored in a computer-readable storage medium, and when the computer program is executed by a processor, the steps of the method embodiments may be implemented. Wherein the computer program comprises computer program code, which may be in the form of source code, object code, an executable file or some intermediate form, etc.
The present invention is not limited to the above-described embodiments, and various modifications and variations of the present invention are intended to be included within the scope of the claims and the equivalent technology of the present invention if they do not depart from the spirit and scope of the present invention.
Claims (10)
1. A cross-domain emotion analysis method based on a mutual learning network is characterized by comprising the following steps:
acquiring a training text data set, wherein the training text data set comprises a text data set of a source field and a text data set of a target field, the text data set comprises a plurality of text data, the text data comprises a plurality of sentences, and the sentences comprise a plurality of words;
inputting the training text data set into a preset word embedding model to obtain a word embedding vector set, wherein the word embedding vector set comprises word embedding vector representations corresponding to text data of a plurality of source fields and word embedding vector representations corresponding to text data of a plurality of target fields;
inputting the word embedding vector set into a preset mutual learning network, constructing a loss function corresponding to the mutual learning network, performing optimization training, and obtaining a target mutual learning network, wherein the mutual learning network comprises two groups of mutual learning migration channels;
and responding to an analysis instruction, acquiring text data to be analyzed, inputting the text data to be analyzed into the target mutual learning network, and acquiring an emotion analysis result output by the target mutual learning network.
2. The cross-domain emotion analysis method based on mutual learning network as claimed in claim 1, wherein:
the two groups of mutual learning migration channels respectively comprise a feature extraction module, an emotion classification module, a field difference learning module and a label detection module;
inputting the word embedding vector set into a preset mutual learning network, and constructing a loss function corresponding to the mutual learning network, including the steps of:
respectively inputting the word embedding vector set to the feature extraction modules in the two groups of mutual learning migration channels, and acquiring text feature representations corresponding to the text data of the plurality of source fields and the text feature representations corresponding to the text data of the plurality of target fields, which are output by the feature extraction modules of the two groups of mutual learning migration channels;
respectively inputting text feature representations corresponding to the text data of the source fields and text feature representations corresponding to the text data of the target fields into emotion classification modules of the two groups of mutual learning migration channels, obtaining emotion feature representations corresponding to the text data of the source fields and emotion feature representations corresponding to the text data of the target fields of the two groups of mutual learning migration channels, and constructing first loss functions of the two groups of mutual learning migration channels according to the emotion feature representations corresponding to the text data of the source fields;
respectively inputting text feature representations corresponding to the text data of the target domains into label detection modules of the two groups of mutual learning migration channels, acquiring label feature representations corresponding to the text data of the target domains of the two groups of mutual learning migration channels, inputting label feature representations corresponding to the text data of the target domains of the mutual learning migration channels into an emotion classification module of the other group of mutual learning migration channels, and constructing a second loss function of the two groups of mutual learning migration channels according to emotion feature representations corresponding to the text data of the source domains and label feature representations corresponding to the text data of the target domains;
inputting the text feature representations corresponding to the text data of the source fields and the text feature representations corresponding to the text data of the target fields into a field difference learning module of the quantity group mutual learning migration channels, and constructing a third loss function of the two groups of mutual learning migration channels;
respectively constructing total loss functions corresponding to the two groups of mutual learning migration channels according to the first loss function, the first loss function and the third loss function of the two groups of mutual learning migration channels, and performing optimization training according to the total loss functions corresponding to the two groups of mutual learning migration channels to obtain the target mutual learning network.
3. The cross-domain emotion analysis method based on mutual learning network as claimed in claim 2, wherein:
the feature extraction module comprises a semantic feature extraction module and a syntactic feature extraction module which are connected in sequence;
the step of inputting the word embedding vector set into the feature extraction modules in the two groups of mutual learning migration channels respectively to obtain text feature representations corresponding to the text data of the plurality of source fields and the text feature representations corresponding to the text data of the plurality of target fields, which are output by the feature extraction modules in the two groups of mutual learning migration channels, includes the steps of:
respectively inputting the word embedding vector set to semantic feature extraction modules in the corresponding feature extraction modules, and respectively acquiring semantic feature representations corresponding to the text data of the plurality of source fields and the text data of the plurality of target fields, which are output by the semantic feature extraction modules of the two groups of mutual learning migration channels;
respectively inputting the word embedding vector set to a syntactic feature extraction module in the corresponding feature extraction module, and respectively acquiring syntactic feature representations corresponding to the text data of the plurality of source fields and the syntactic feature representations corresponding to the text data of the plurality of target domains, which are output by the syntactic feature extraction modules of the two groups of mutual learning migration channels;
and respectively splicing the semantic feature representation and the syntactic feature representation corresponding to the text data of the source fields and the semantic feature representation and the syntactic feature representation corresponding to the text data of the target fields, and acquiring the text feature representation corresponding to the text data of the source fields and the text feature representation corresponding to the text data of the target fields, which are output by the syntactic feature extraction modules of the two groups of mutual learning migration channels.
4. The cross-domain emotion analysis method based on mutual learning network as claimed in claim 3, wherein:
the semantic feature extraction module comprises a first bidirectional gated loop unit, a soft attention unit, a second bidirectional gated loop unit and a convolution attention unit, wherein the convolution attention unit comprises a plurality of convolution layers;
the method comprises the following steps of respectively inputting the word embedding vector set into semantic feature extraction modules in the corresponding feature extraction modules, respectively obtaining semantic feature representations corresponding to the text data of the plurality of source fields and the text data of the plurality of target fields, which are output by the semantic feature extraction modules of the two groups of mutual learning migration channels, and respectively:
inputting the word embedding vector set into a first bidirectional gating circulating unit in the corresponding semantic feature extraction module, respectively encoding the word embedding vector set, and acquiring hidden layer feature representations corresponding to the text data of the plurality of source fields and the text data of the plurality of target fields, which are output by the first bidirectional gating circulating unit of the semantic feature extraction modules of the two groups of mutual learning migration channels;
inputting hidden layer feature representations corresponding to the text data of the source domains and hidden layer feature representations corresponding to the text data of the target domains into corresponding soft attention units in the semantic feature extraction module, acquiring corresponding attention weight parameters and attention weight parameters corresponding to the text data of the target domains according to a preset attention weight parameter calculation algorithm, and respectively acquiring sentence feature representations corresponding to the text data of the source domains and sentence feature representations corresponding to the text data of the target domains, which are output by the soft attention units of the two groups of mutual learning migration channels, according to a preset sentence feature representation calculation algorithm, wherein the attention weight parameter calculation algorithm is as follows:
in the formula (I), the compound is shown in the specification,
for the purpose of the attention weight parameter,
for a preset first trainable network parameter,
for a preset second trainable network parameter,
is a preset third trainable network parameter;
the sentence characteristic representation calculation algorithm is as follows:
in the formula (I), the compound is shown in the specification,
characterizing the sentence;
respectively inputting sentence characteristic representations corresponding to the text data of the source fields and sentence characteristic representations corresponding to the text data of the target fields into second bidirectional gating circulation units in the corresponding semantic characteristic extraction modules for coding, and respectively acquiring the sentence characteristic representations corresponding to the text data of the source fields and the sentence characteristic representations corresponding to the text data of the target fields after coding, which are output by the second bidirectional gating circulation units of the two groups of mutual learning migration channels;
inputting the sentence characteristic representations corresponding to the text data of the plurality of source fields after the encoding processing and the sentence characteristic representations corresponding to the text data of the plurality of target fields after the encoding processing into corresponding convolution attention units in the semantic characteristic extraction module, respectively performing weighting processing on the sentence characteristic representations corresponding to the text data of the plurality of source fields after the encoding processing and the sentence characteristic representations corresponding to the text data of the plurality of target fields after the encoding processing according to a preset attention algorithm, and acquiring the semantic characteristic representations corresponding to the text data of the plurality of source fields and the semantic characteristic representations corresponding to the text data of the plurality of target fields, wherein the attention algorithm is as follows:
in the formula (I), the compound is shown in the specification,qfor the number of windows in the convolution attention unit,mfor the number of sentences in the text data,tis the first in the convolution attention unittA convolution filter for the received signal and a convolution filter for the received signal,
in order to represent the characteristics of the convolution,
in order to be a non-linear activation function,
denotes the firsttThe function operation of each convolution filter is carried out,
which represents a convolution operation, the operation of the convolution,
is a function of the pre-set convolution coefficients,
is as followskSemantic feature representations corresponding to the text data of the individual target domains.
5. The cross-domain emotion analysis method based on mutual learning network as claimed in claim 3, wherein:
the syntactic characteristic extraction module comprises a bidirectional gating circulation unit, a multilayer drawing attention network unit and a convolution unit, wherein the multilayer drawing attention network unit comprises a plurality of drawing attention network layers;
the method comprises the following steps of inputting the word embedding vector set into corresponding syntactic feature extraction modules in the feature extraction modules respectively, and acquiring syntactic feature representations corresponding to the text data of the source fields and the text data of the target fields, which are output by the syntactic feature extraction modules of the two groups of mutual learning migration channels, respectively, and the syntactic feature representations corresponding to the text data of the target fields, and comprises the following steps:
inputting the word embedding vector set to a bidirectional gating circulating unit in the corresponding syntactic feature extraction module, and respectively acquiring context hidden layer feature representations corresponding to the text data of the plurality of source fields and the text data of the plurality of target fields, which are output by the bidirectional gating circulating unit of the syntactic feature extraction module of the two groups of mutual learning migration channels;
inputting the context hidden layer feature representations corresponding to the text data of the source fields and the context hidden layer feature representations corresponding to the text data of the target fields into corresponding multilayer drawing attention network units in the syntactic feature extraction module, taking the context hidden layer feature representations as state vectors of a first drawing attention network layer of the multilayer drawing attention network units, and respectively acquiring the state vectors corresponding to nodes of the drawing attention network layers of the two groups of mutual learning migration channels according to a preset state vector calculation algorithm, wherein the state vector calculation algorithm is as follows:
in the formula (I), the compound is shown in the specification,
is a non-linear activation function;
is at the firstlPersonal attention network layer, the firstxA node and ayRaw attention score data between individual nodes,
for a pre-set trainable weight vector, | | | is expressed as a concatenation operation,
for a pre-set trainable weight matrix,
is as followslFirst of the layerxThe state vector corresponding to each node is calculated,
after normalization processing, in the second steplPersonal attention network layer, the firstxA node and ayAn attention weight parameter between the individual nodes,
in order to be a non-linear activation function,
for all and nodesxA set of adjacent nodes;
and inputting the state vectors corresponding to the nodes of the plurality of graph attention network layers of the two groups of mutual learning migration channels into the corresponding convolution units in the syntactic feature extraction module for feature extraction, and acquiring syntactic feature representations corresponding to the text data of the plurality of source fields and syntactic feature representations corresponding to the text data of the plurality of target fields.
6. The cross-domain emotion analysis method based on mutual learning network of claim 2, wherein the text feature representations corresponding to the text data of the source domains and the text feature representations corresponding to the text data of the target domains are respectively input to emotion classification modules of the two groups of mutual learning migration channels, emotion feature representations corresponding to the text data of the source domains of the two groups of mutual learning migration channels and emotion feature representations corresponding to the text data of the target domains are obtained, and the first loss functions of the two groups of mutual learning migration channels are constructed according to the emotion feature representations corresponding to the text data of the source domains, including:
according to the text feature representations corresponding to the text data of the source fields, the text feature representations corresponding to the text data of the target fields and the emotion classification modules of the two groups of mutual learning migration channels, a preset emotion feature representation calculation algorithm is used for respectively obtaining emotion feature representations corresponding to the text data of the source fields of the two groups of mutual learning migration channels and emotion feature representations corresponding to the text data of the target fields, wherein the emotion feature representation calculation algorithm is as follows:
in the formula (I), the compound is shown in the specification,
representing corresponding emotional characteristics of the text data of the source field,
is the emotion characteristic representation corresponding to the text data of the target domain,
in order to be a function of the normalization,
for a preset fourth trainable network parameter,
is a preset fifth trainable network parameter;
obtaining label data corresponding to the text data of the source fields, and constructing a first loss function of the two groups of mutual learning migration channels according to emotion feature representations and the label data corresponding to the text data of the source fields, wherein the first loss function is as follows:
in the formula (I), the compound is shown in the specification,
for the purpose of said first loss function,
the number of text data of the source domain is set for the text data of the source domain,
is as followskAnd label data corresponding to the text data of the source field.
7. The cross-domain emotion analysis method based on mutual learning network as claimed in claim 6, characterized in that the text feature representations corresponding to the text data of the target domains are respectively input into the label detection modules of the two groups of mutual learning migration channels, acquiring label feature representations corresponding to the text data of a plurality of target domains of the two groups of mutual learning migration channels, and label feature representations corresponding to the text data of a plurality of target domains of one group of the mutual learning migration channels are input into an emotion classification module of another group of the mutual learning migration channels, constructing a second loss function of the two groups of mutual learning migration channels according to the emotional characteristic representations corresponding to the text data of the plurality of source fields and the label characteristic representations corresponding to the text data of the plurality of target fields, comprising the following steps:
according to the text feature representation corresponding to the text data of the target domains and a preset label feature representation calculation algorithm in the label detection modules of the two groups of mutual learning migration channels, obtaining label feature representations corresponding to the text data of the target domains of the two groups of mutual learning migration channels, wherein the label feature representation calculation algorithm is as follows:
in the formula (I), the compound is shown in the specification,
label feature representations corresponding to the text data of the several target fields,
for a preset sixth trainable network parameter,
is a preset seventh trainable network parameter;
inputting label feature representations corresponding to text data of a plurality of target domains of one group of mutual learning migration channels into an emotion classification module of another group of mutual learning migration channels, and constructing second loss functions of the two groups of mutual learning migration channels according to emotion feature representations corresponding to the text data of the plurality of source domains and label feature representations corresponding to the text data of the plurality of target domains, wherein the second loss functions are as follows:
in the formula (I), the compound is shown in the specification,
、
respectively learning second loss functions corresponding to the migration channels for the two groups
、
The emotion feature representations corresponding to the text data of the target domains corresponding to the two groups of mutual learning migration channels are respectively displayed,
、
label feature representations corresponding to the text data of the target domains corresponding to the two groups of mutual learning migration channels respectively,
is a preset divergence coefficient.
8. The cross-domain emotion analysis method based on mutual learning network as claimed in claim 6, wherein the step of inputting the text feature representations corresponding to the text data of the source domains and the text feature representations corresponding to the text data of the target domains into the domain difference learning module of the quantity group mutual learning migration channels to construct the third loss function of the two groups of mutual learning migration channels comprises the steps of:
combining the text characteristic representations corresponding to the text data of the source fields and the text characteristic representations corresponding to the text data of the target fields respectively to obtain the text characteristic representation corresponding to the text data set of the source fields and the text characteristic representation corresponding to the text data set of the target fields;
inputting the text feature representation corresponding to the text data set of the source field and the text feature representation corresponding to the text data set of the target field into a field difference learning module of the quantity group mutual learning migration channel, and constructing a bulldozer distance function, wherein the bulldozer distance function is as follows:
in the formula (I), the compound is shown in the specification,
as a function of the bulldozer distance,
the number of text data of the source domain is set to the text data of the source domain,
is the number of text data of the target field in the text data set of the target field,
the source domain is represented by a field of view,
a target domain is represented by a target field,
for the learning function in the domain difference learning module,
a text feature representation corresponding to the text dataset of the source domain,
representing the text characteristic corresponding to the text data set of the target domain;
constructing a gradient penalty function according to preset learning training parameters and balance coefficients, and acquiring a third loss function of the two groups of field difference learning modules of the mutual learning migration channels according to the bulldozer distance function and the gradient penalty function, wherein the gradient penalty function is as follows:
in the formula (I), the compound is shown in the specification,
is a gradient penalty characteristic representation
And
the serial connection of random points on the feature space;
is the equilibrium coefficient.
9. The method for cross-domain emotion analysis based on mutual learning network as claimed in claim 2, wherein the step of inputting the text data to be analyzed into the target mutual learning network and obtaining emotion analysis results output by the target mutual learning network comprises the steps of:
inputting the text data to be analyzed into the target mutual learning network, obtaining text feature representation of the text data to be analyzed, and obtaining emotion polarity vectors of the text data to be analyzed according to the text feature representation of the text data to be analyzed and a preset emotion polarity vector calculation algorithm, wherein the emotion polarity vector calculation algorithm is as follows:
in the formula (I), the compound is shown in the specification,pfor the said vector of the emotion polarity,
in order to be a function of the normalization,
the parameters are updated for the purposes of the weights,
the bias term is updated for the weight(s),
representing the text characteristics of the text data to be analyzed;
and acquiring the emotion polarity corresponding to the dimensionality with the maximum probability according to the emotion classification polarity probability distribution vector, and using the emotion polarity as an emotion analysis result output by the target mutual learning network.
10. A cross-domain emotion analysis device based on a mutual learning network is characterized by comprising:
the training text data set acquisition module is used for acquiring a training text data set, wherein the training text data set comprises text data of a plurality of source fields and text data of a plurality of target fields, the text data comprises a plurality of sentences, and the sentences comprise a plurality of words;
a word embedding vector representation obtaining module, configured to input the training text data set into a preset word embedding model, and obtain a word embedding vector set, where the word embedding vector set includes word embedding vector representations corresponding to text data of a plurality of source fields and word embedding vector representations corresponding to text data of a plurality of target fields;
the network training module is used for inputting the word embedding vector set into a preset mutual learning network, constructing a loss function corresponding to the mutual learning network, performing optimization training and obtaining a target mutual learning network, wherein the mutual learning network comprises two groups of mutual learning migration channels;
and the text data analysis module is used for responding to an analysis instruction, acquiring text data to be analyzed, inputting the text data to be analyzed into the target mutual learning network, and acquiring an emotion analysis result output by the target mutual learning network.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210954299.9A CN115033700A (en) | 2022-08-10 | 2022-08-10 | Cross-domain emotion analysis method, device and equipment based on mutual learning network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210954299.9A CN115033700A (en) | 2022-08-10 | 2022-08-10 | Cross-domain emotion analysis method, device and equipment based on mutual learning network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115033700A true CN115033700A (en) | 2022-09-09 |
Family
ID=83130758
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210954299.9A Pending CN115033700A (en) | 2022-08-10 | 2022-08-10 | Cross-domain emotion analysis method, device and equipment based on mutual learning network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115033700A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115712726A (en) * | 2022-11-08 | 2023-02-24 | 华南师范大学 | Emotion analysis method, device and equipment based on bigram embedding |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113065344A (en) * | 2021-03-24 | 2021-07-02 | 大连理工大学 | Cross-corpus emotion recognition method based on transfer learning and attention mechanism |
| CN114331123A (en) * | 2021-12-28 | 2022-04-12 | 重庆邮电大学 | Teaching evaluation emotion analysis method integrating cognitive migration |
-
2022
- 2022-08-10 CN CN202210954299.9A patent/CN115033700A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113065344A (en) * | 2021-03-24 | 2021-07-02 | 大连理工大学 | Cross-corpus emotion recognition method based on transfer learning and attention mechanism |
| CN114331123A (en) * | 2021-12-28 | 2022-04-12 | 重庆邮电大学 | Teaching evaluation emotion analysis method integrating cognitive migration |
Non-Patent Citations (3)
| Title |
|---|
| CHI YANG ET AL.: "Dual-Channel Domain Adaptation Model", 《IEEE/WIC/ACM INTERNATIONAL CONFERENCE ON WEB INTELLIGENCE AND INTELLIGENT AGENT TECHNOLOGY》 * |
| JIAN SHEN ET AL.: "Wasserstein Distance Guided Representation Learning for Domain Adaptation", 《THE THIRTY-SECOND AAAI CONFERENCE ON ARTIFICIAL INTELLIGENCE》 * |
| QIANMING XUE ET AL.: "Improving Domain-Adapted Sentiment Classification by Deep Adversarial Mutual Learning", 《THE THIRTY-FOURTH AAAI CONFERENCE ON ARTIFICIAL INTELLIGENCE》 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115712726A (en) * | 2022-11-08 | 2023-02-24 | 华南师范大学 | Emotion analysis method, device and equipment based on bigram embedding |
| CN115712726B (en) * | 2022-11-08 | 2023-09-12 | 华南师范大学 | Emotion analysis method, device and equipment based on double word embedding |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112819023B (en) | Sample set acquisition method, device, computer equipment and storage medium | |
| CN109559300A (en) | Image processing method, electronic equipment and computer readable storage medium | |
| CN111950596A (en) | Training method for neural network and related equipment | |
| CN114676704B (en) | Sentence emotion analysis method, device and equipment and storage medium | |
| JP2018022496A (en) | Method and equipment for creating training data to be used for natural language processing device | |
| CN115587597B (en) | Sentiment analysis method and device of aspect words based on clause-level relational graph | |
| CN111368656A (en) | Video content description method and video content description device | |
| McCormack et al. | Understanding aesthetic evaluation using deep learning | |
| CN116151263B (en) | Multi-mode named entity recognition method, device, equipment and storage medium | |
| CN115168592B (en) | Statement emotion analysis method, device and equipment based on aspect categories | |
| Park et al. | Neurocartography: Scalable automatic visual summarization of concepts in deep neural networks | |
| CN112749737A (en) | Image classification method and device, electronic equipment and storage medium | |
| CN115879508A (en) | Data processing method and related device | |
| CN115033700A (en) | Cross-domain emotion analysis method, device and equipment based on mutual learning network | |
| Qayyum et al. | Ios mobile application for food and location image prediction using convolutional neural networks | |
| CN111445545B (en) | Text transfer mapping method and device, storage medium and electronic equipment | |
| CN115905518B (en) | Emotion classification method, device, equipment and storage medium based on knowledge graph | |
| CN115906863B (en) | Emotion analysis method, device, equipment and storage medium based on contrast learning | |
| CN114547312B (en) | Emotional analysis method, device and equipment based on common sense knowledge graph | |
| CN115659951A (en) | Statement emotion analysis method, device and equipment based on label embedding | |
| CN112132269B (en) | Model processing method, device, equipment and storage medium | |
| CN115906861A (en) | Statement emotion analysis method and device based on interaction aspect information fusion | |
| CN115204171A (en) | Document-level event extraction method and system based on hypergraph neural network | |
| CN113821610A (en) | Information matching method, device, equipment and storage medium | |
| CN113627522A (en) | Image classification method, device and equipment based on relational network and 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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20220909 |