CN116244429A - Microblog Sentiment Analysis Method Based on Multi-level Feature Interaction Fusion Guided by Social Relations - Google Patents

Abstract

The invention relates to a microblog emotion analysis method with multi-level feature interaction fusion guided by social relations, and belongs to the field of natural language processing. The invention comprises the following steps: extracting word-level and sentence-level characteristics and relation characteristics of the microblog text by using BERT and LINE respectively; fusing word-level features by using CNNs with double channels and multiple convolution kernels; the word-level features and sentence-level features are guided by using the relation features to perform feature interaction, interaction noise is eliminated, and word-level and sentence feature representation after interaction is obtained; constructing an attention feature fusion network to fuse word-level features after interaction and sentence-level features after interaction for one time; the dynamic weighting coefficients are used for carrying out weight distribution on the primary fusion features, the interactive word-level features and the interactive sentence-level features; and constructing an interactive fusion network, carrying out secondary fusion on the weighted characteristics, and then carrying out emotion classification. According to the method, the microblog relation network is filtered by considering a plurality of factors in the social behavior of the user, and the probability of emotion consistency among the microblogs in the microblog relation network is improved.

Description

Microblog sentiment analysis method based on multi-level feature interaction fusion guided by social relationship

Technical Field

The present invention belongs to the field of natural language processing and relates to a microblog sentiment analysis method based on multi-level feature interaction fusion guided by social relationships.

Background Art

In recent years, Twitter and Sina Weibo have become popular platforms for users to post and spread information on Weibo. Analyzing the sentiment of Weibo on large social platforms can help governments or e-commerce companies perceive the public's views on various topics (such as political events, celebrities, daily life, etc.), and it has a wide range of applications in academic and industrial fields. Weibo sentiment analysis aims to quickly judge the sentiment of massive data, reduce human input, and provide decision makers with timely feedback information for quick decision-making.

The basic idea of the microblog sentiment analysis method is to train a sentiment classifier on a dataset that has been manually labeled with sentiment, and then use the classifier as a sentiment analysis model to identify the sentiment of unlabeled microblogs. It is divided into two steps: preprocessing of microblog text data and establishment of a sentiment analysis model. In terms of preprocessing, after removing special symbols and stop words, each sentence of the microblog text is converted into a corresponding word format through a word segmentation tool. In terms of establishing a sentiment analysis model, a neural network model is constructed to solve the mapping function from microblog text in word format to sentiment features. However, due to the use of new network words and special symbols in microblogs, there is a lack of words in microblogs, which leads to a sparse vocabulary problem. In order to solve this problem, researchers have shifted their attention to the interactive activities between microblog texts (i.e., interactive behaviors such as following, liking, and forwarding on social platforms). They find related microblogs through interactive behaviors and expand the target microblogs to solve the problem of sparse vocabulary.

Undoubtedly, the introduction of interactive activity information can solve the problem of low sentiment recognition rate caused by data noise in Weibo to a certain extent. The early method of introducing interactive activities into Weibo sentiment analysis was to use user attention relationships to build a relationship network between Weibo, so that the relationship Weibo can expand the target Weibo. However, as the number of Weibo posts by users increases, too much data noise is brought to the target Weibo. To address this problem, current research methods take into account individual factors such as topic features, user similarity, and Weibo similarity in the construction of Weibo relationship networks to filter data noise. However, the current existing research methods have the following defects: (1) They do not consider multiple factors to filter the Weibo relationship network to increase the probability of emotional consistency of the relationship network; (2) They only use the target Weibo as an expansion feature to solve the data sparsity problem, and do not allow the Weibo relationship network to participate in guiding the interaction between Weibo. Therefore, the existing technology is not sufficient for the research on Weibo sentiment analysis combined with social relationships.

Summary of the invention

In view of this, the purpose of the present invention is to provide a microblog sentiment analysis method with multi-level feature interaction fusion guided by social relations.

In order to achieve the above object, the present invention provides the following technical solutions:

A social relationship-guided multi-level feature interactive fusion microblog sentiment analysis method, the method comprising the following steps:

Step 1: Extraction of Weibo text and user interaction features. Preprocess Weibo text and user social behavior data, use BERT pre-trained models of different dimensions to extract word-level and sentence-level features of Weibo text, and use LINE to extract relational features of social information to represent them in a tensor format that can be recognized by computers.

Furthermore, the specific process of step 1 includes:

(l为句子的长度)作为词级的表示特征。将BERT模型输出的最后一个隐藏层的[CLS]特征

作为句级的表示特征。则经过BERT编码器，输出的文本特征可以表示为：First, for the microblog text s _i , the features of the state of the last hidden layer in the BERT model are

(l is the length of the sentence) as the word-level representation feature. The [CLS] feature of the last hidden layer output by the BERT model

As a sentence-level representation feature, after the BERT encoder, the output text feature can be expressed as:

表示第1维度的BERT模型输出的词级特征，

表示第1种维度BERT模型输出的句级特征，Linear(X,y)的意义在于给特征X乘上过度的可训练矩阵，使输出的特征维度映射为y维。最终可以得到所有微博文本的特征表示：Among them, d ₀ and d ₁ are the dimensions of the two BERT pre-training models.

Represents the word-level features output by the BERT model in the first dimension,

It represents the sentence-level features output by the BERT model in the first dimension. The meaning of Linear(X,y) is to multiply feature X by an excessive trainable matrix so that the output feature dimension is mapped to the y dimension. Finally, the feature representation of all Weibo texts can be obtained:

用户发布微博的提及标签

用户发布微博的主题标签

三个方面构建微博间的关系权重网络。对于依靠主题标签(#)建立的微博关系的规则为：两个微博处于同一主题下，关系权重表示为两个微博共享相同主题的个数。对于依靠提及标签(@)建立的微微博关系的规则为：(1)微博A提及用户B后，用户B发布微博B，则微博A和微博B间有一条权重为1的边；(2)两个微博有相同提及的用户，权重为提及相同用户的个数。对于依靠用户关注建立的微博关系的规则为：(1)用户之间有关注关系且用户发布微博，则这些微博之间有一条权重为1的边；(2)同一用户发布的微博之间有一条权重为1的边。最终可以构建一个微博关系权重网络E(i,j,w)。Then, the user interaction features are extracted. The user interaction feature extraction is divided into two steps: the construction of the Weibo relationship weight network and the embedding of the Weibo relationship weight network. The Weibo relationship weight network is constructed.

The mention tags of the user's microblog

Theme tags for users to post on Weibo

The relationship weight network between microblogs is constructed from three aspects. The rule for the microblog relationship established by the topic tag (#) is: if two microblogs are under the same topic, the relationship weight is expressed as the number of topics shared by the two microblogs. The rule for the microblog relationship established by the mention tag (@) is: (1) after microblog A mentions user B, user B posts microblog B, there is an edge with a weight of 1 between microblog A and microblog B; (2) if two microblogs have the same mentioned user, the weight is the number of times the same user is mentioned. The rule for the microblog relationship established by user following is: (1) if there is a following relationship between users and users post microblogs, there is an edge with a weight of 1 between these microblogs; (2) there is an edge with a weight of 1 between microblogs posted by the same user. Finally, a microblog relationship weight network E(i,j,w) can be constructed.

Finally, embedding of the Weibo relation weight network. LINE is used to embed the nodes in a weight network as corresponding low-dimensional feature vectors. For the Weibo relation weight network E(i,j,w) established in the first step, Weibo nodes v _i and v _j , the embedding representation of each Weibo node in the weight network can be obtained by minimizing the embedding loss function O.

表示微博i的低维嵌入向量，

表示当

视为上下文的向量。生成微博关系嵌入表示为

微博数量为m，微博节点嵌入维度是d_r。其中，没有参与到社交互动的微博节点填充为相同维度的噪声向量

得到m个微博节点的嵌入矩阵表示为

Among them, (i,j,wi _,j ) indicates that there is an edge with weight wi _,j between Weibo i and Weibo j.

represents the low-dimensional embedding vector of Weibo i,

Indicates when

The generated microblog relationship embedding is represented as

The number of microblogs is m, and the embedding dimension of microblog nodes is d _r . Among them, microblog nodes that do not participate in social interactions are filled with noise vectors of the same dimension

The embedding matrix of m microblog nodes is expressed as

Step 2: Fusion of word-level features: It is planned to use a dual-channel CNN with multiple convolution kernels of different sizes to fuse word-level features.

Further, the specific process of step 2 includes:

采用CNN提取到的词级特征输出

表示为：First, the word-level feature vector _extracted from the microblog text si

Output of word-level features extracted by CNN

It is expressed as:

y＝{y ¹ ,y ² ,...,y ^L }

为两个BERT输出到卷积双通道的堆叠，W^l和b^l分别表示第l个卷积核的权重矩阵和偏置，

表示卷积操作；

表示在一个包含h个词的窗口中，第l个卷积核的输出。接着，通过最大池化层找出文本中的重要部分p并在最后一维度连接所有特征。得到最终的词级特征

表示为：Where ^yi represents the output of the i-th convolution kernel,

is a stack of two BERT outputs to the convolution dual channel, W ^l and b ^l represent the weight matrix and bias of the lth convolution kernel respectively,

Represents the convolution operation;

represents the output of the lth convolution kernel in a window containing h words. Next, the important part p in the text is found through the maximum pooling layer and all features are connected in the last dimension. The final word-level feature is obtained

It is expressed as:

p=cat([maxpool(y ¹ ),maxpool(y ² ),...,maxpool(y ^L )])

表示为Then, in order to unify the dimensions, the word-level feature p is mapped to d _r dimensions, that is,

Expressed as

_ps ＝Linear(p, _dr )

也映射为d_r维，即

表示为：Finally, sentence-level features are extracted. The sentence-level features of the first BERT are used as overall sentence features.

It is also mapped to d _r dimension, that is,

It is expressed as:

Step 3: Interaction of word-level, sentence-level, and relational features. It is planned to use the relational features in step 1 to guide the word-level features in step 2 and the sentence-level features in step 1 to interact, and obtain the word-level features and sentence-level features after interaction.

Further, the specific process of step 3 includes:

The first is the guided interaction of relational features, including the construction of microblog similarity matrix, and guided interaction at word level and sentence level.

对于第i个微博节点嵌入

得到微博相似性矩阵M为：Then, the microblog similarity matrix is constructed. The microblog similarity matrix is constructed using the relationship features and is defined as the normalized value of the similarity between two node vectors. The embedding matrix of m microblog nodes is expressed as

For the i-th microblog node embedding

The microblog similarity matrix M is obtained as:

h＝Tanh(sum _r (M _r -I ₍₁₎ *v _noise )*sum _r (M _r -I*v _noise ))

M _corr =ReLU(Tanh((h·h ^T -I)*10))

Where || _vi || _F is the Frobenius norm of vector _vi , · is matrix multiplication, * is element multiplication, and sum _r () represents summing the matrix by row. I ₍₁₎ is a matrix of all 1s, and I is the identity matrix. M _corr is a correction matrix, which is mainly used to correct the microblog text similarity noise caused by the filling vector v _noise of the randomly filled microblog nodes. Tanh and ReLU are activation functions, as follows:

ReLU(x)=max(0,x)

交互后句级特征

输出分别为：Finally, the guided interaction between word level and sentence level. For m microblogs, the word level features after interaction

Post-interaction sentence-level features

The outputs are:

Step 4: First fusion of features. Take the relational features of step 1 as Q (query vector), the sentence-level features of step 3 as K (key-value vector), and the word-level features of step 3 as V (value vector) as features, and build an attention fusion network to fuse the three features for the first time to obtain a fusion feature. Specifically, take the microblog relational feature _Mr as the guided query (Q), the sentence-level feature _hcls after interaction as the key value (K), and the word-level feature _hs as the value (V), and input them into the self-attention mechanism to obtain the guided microblog text’s first fusion feature:

h _att = SelfAtt (Q = M _r , K = h _s , V = h _cls )

Where d _k is the dimension of the key-value vector K. Q, K, V are the corresponding query, key-value, and value vectors.

Step 5: Second fusion of features. Dynamic weighting coefficients are used to assign weights to the first fusion features of step 4, the word-level features after interaction of step 3, and the sentence-level features after interaction. An interactive fusion network is constructed to perform a second fusion of the weighted features.

Further, the specific process of step 5 includes:

First, dynamic weighting is performed. The first fusion feature h _att , the word-level feature after interaction h _cls , and the sentence-level feature after interaction h _s are stacked and dynamically weighted. The weighted feature is:

h _stack =Stack([h _s ,h _att ,h _cls ])

h _c =(h _stack ) ^T *[a ₁ ,a ₂ ,a ₃ ]

为堆叠后的特征，[a₁,a₂,a₃]为3个特征的动态加权系数。in,

is the stacked feature, and [a ₁ ,a ₂ ,a ₃ ] is the dynamic weighting coefficient of the three features.

Next, the interactive fusion network. Similar to the word-level feature fusion network, CNN is used to perform a second fusion of the weighted features, and the fused feature h is:

h＝{h ¹ ,h ² ,...,h ^L }

Finally, the interactive features are obtained through average pooling and concatenated in the last dimension. The fused features are:

h _f =cat([avgpool(h ¹ ),avgpool(h ² ),...,avgpool(h ^L )])

Step 6: Weibo sentiment classification. Construct a Softmax sentiment classifier to classify Weibo sentiment. Use the cross entropy loss function as the training loss function and use the back propagation algorithm to train the model to obtain a Weibo sentiment analysis model.

Further, the specific process of step 6 includes:

First, build a Softmax classifier to complete the classification of text sentiment:

Among them, num_class is the sentiment category corresponding to the microblog text.

Finally, the back propagation algorithm is used to train the model, and the cross entropy loss function is used as the loss function in the training process. The model is optimized by optimizing the loss function. The expression is:

分别为样本的真实样本概率分布和预测样本概率分布。λ为L2正则化的系数。Among them, J(w,b) is the loss of the whole sample, m is the number of samples, y ⁽ⁱ⁾ ,

are the true sample probability distribution and predicted sample probability distribution of the sample respectively. λ is the coefficient of L2 regularization.

The beneficial effects of the present invention are:

1) Aiming at the low probability of sentiment consistency in the Weibo relationship network constructed by the existing Weibo sentiment analysis method combined with social relationships, multiple factors are considered to filter the Weibo relationship network to improve the probability of sentiment consistency in the relationship network.

2) The existing methods only use the target Weibo as an extended feature to solve the data sparsity problem, and do not allow the Weibo relationship network to participate in guiding the interaction between Weibo. Therefore, a sentiment analysis network is extracted that uses social relationships to guide Weibo text interaction.

3) For users who do not participate in social activities on the social platform, noise is generated when introducing social relationships, and noise is eliminated.

4) Use the dual-channel BERT pre-trained model to jointly extract word-level and sentence-level features of microblog texts, enriching the text features. At the same time, design a weighted fusion network to fuse text features to avoid single-channel misjudgment.

Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description.

BRIEF DESCRIPTION OF THE DRAWINGS

In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

FIG1 is a flow chart of a BERT-based microblog sentiment analysis network method guided by social relationships of the present invention;

FIG2 is a model diagram of a microblog sentiment analysis network system based on BERT under the guidance of social relationships of the present invention.

DETAILED DESCRIPTION

The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and the features in the embodiments can be combined with each other without conflict.

Among them, the drawings are only used for illustrative explanations, and they only represent schematic diagrams rather than actual pictures, and should not be understood as limitations on the present invention. In order to better illustrate the embodiments of the present invention, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if the terms "upper", "lower", "left", "right", "front", "back" and the like indicate directions or positional relationships, they are based on the directions or positional relationships shown in the drawings, which are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific direction, be constructed and operated in a specific direction. Therefore, the terms describing the positional relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

As shown in Figure 1, the present invention provides a social relationship-guided multi-level feature interaction fusion microblog sentiment analysis method. The implementation scenario is that for a data set containing microblog texts and corresponding sentiment and social information, a network model for sentiment analysis is trained using the method mentioned above, so that the model completes the sentiment classification of microblog texts. The specific implementation steps are as follows:

Step 1: Use the BERT pre-trained model and LINE to pre-process Weibo text and social information, that is, use vectors to represent Weibo text and social information.

(l为句子的长度)和句级特征

而本文中使用d₀、d₁两种不同维度的BERT来共同提取特征，如下：First, the BERT pre-trained model is a deep model trained by Google on a large number of unlabeled datasets. It can represent the input microblog text in the corresponding vector format, that is, the word-level and sentence-level features of the microblog text. Specifically, for the microblog text s _i , after BERT, the word-level features of s _i can be obtained.

(l is the length of the sentence) and sentence-level features

In this paper, BERT with two different dimensions, d ₀ and d _1, is used to extract features together, as follows:

文本中的提及关系

文本中的主题关系

来构建微博关系网络。具体地，对于用户间的关注关系

如果用户之间有关注关系且用户发布微博，则这些微博之间有一条权重为1的边；同一用户发布的微博之间有一条权重为1的边。对文本中的提及关系

微博A提及用户B后，用户B发布微博B，则微博A和微博B间有一条权重为1的边；两个微博有相同提及的用户，权重为提及相同用户的个数。对于文本中的主题关系

两个微博处于同一主题下，则两个微博之间有相同主题的个数权重的边。最终得到一个微博关系权重网络E(i,j,w)。Next, through the following relationship between users

Mentions in text

Thematic relationships in text

To build a microblog relationship network. Specifically, for the attention relationship between users

If there is a follow relationship between users and users post microblogs, there is an edge with a weight of 1 between these microblogs; there is an edge with a weight of 1 between microblogs posted by the same user.

After Weibo A mentions user B, user B posts Weibo B, then there is an edge with a weight of 1 between Weibo A and Weibo B; if two Weibo posts have the same mentioned user, the weight is the number of times the same user is mentioned.

If two microblogs are under the same topic, then there is an edge with the same number of topics between the two microblogs. Finally, a microblog relationship weight network E(i,j,w) is obtained.

Finally, the microblog relationship weight network E(i,j,w) is input into the LINE network to obtain the vector representation of each microblog. Specifically, the vector representation of each microblog node, i.e., the relationship feature, is obtained by optimizing the following function:

Step 2: Use CNN to fuse word-level features, use relational features to guide word-level and sentence-level features, and perform the first fusion.

采用CNN提取到的词级特征输出

再通过最大池化层找出文本中的重要部分p并在最后一维度连接所有特征。得到最终的词级特征

表示为：First, word-level features are fused through CNN. For the word-level feature vector extracted from the microblog text _si

Output of word-level features extracted by CNN

Then use the maximum pooling layer to find the important part p in the text and connect all the features in the last dimension to get the final word-level feature

It is expressed as:

y＝{y ¹ ,y ² ,...,y ^L }

p=cat([maxpool(y ¹ ),maxpool(y ² ),...,maxpool(y ^L )])

表示为：Map the word-level feature p to d _r dimensions, that is

It is expressed as:

_ps ＝Linear(p, _dr )

也映射为d_r维，即

表示为：Sentence Features

It is also mapped to d _r dimension, that is,

It is expressed as:

得到微博相似性矩阵M为：Next, we use relational features to guide word-level and sentence-level features. For the i-th microblog node embedding

The microblog similarity matrix M is obtained as:

h＝Tanh(sum _r (M _r -I ₍₁₎ *v _noise )*sum _r (M _r -I*v _noise ))

M _corr =ReLU(Tanh((h·h ^T -I)*10))

交互后句级特征

输出分别为：For m microblogs, word-level features after interaction

Post-interaction sentence-level features

The outputs are:

Finally, the first fusion is performed. The microblog relationship feature _Mr is used as the guided query (Q), the sentence-level feature _hcls after interaction is used as the key value (K), and the word-level feature _hs is used as the value (V). These are input into the self-attention mechanism to obtain the guided microblog text’s first fusion feature:

h _att = SelfAtt (Q = M _r , K = h _s , V = h _cls )

Step 3: Build a feature fusion network to perform secondary fusion of features and perform sentiment classification on microblogs.

First, the first fusion feature h _att , the word-level feature after interaction h _cls , and the sentence-level feature after interaction h _s are stacked and dynamically weighted. The weighted feature is:

h _stack =Stack([h _s ,h _att ,h _cls ])

h _c =(h _stack ) ^T *[a ₁ ,a ₂ ,a ₃ ]

Next, similar to the word-level feature fusion network, CNN is used to fuse the weighted features for the second time, and the fused feature h is:

h＝{h ¹ ,h ² ,...,h ^L }

The interactive features are obtained through average pooling and concatenated in the last dimension. The fused features are:

h _f =cat([avgpool(h ¹ ),avgpool(h ² ),...,avgpool(h ^L )])

Then, build a Softmax classifier to complete the classification of text sentiment:

Finally, the back propagation algorithm is used to train the model, and the cross entropy loss function is used as the loss function in the training process. The model is optimized by optimizing the loss function. The expression is:

After completing the above steps, you will eventually get a model that can be used for Weibo sentiment classification. After saving the model, you can use it. Given a Weibo text, you can get the sentiment tendency of the model by inputting it into the model.

FIG2 is a system model diagram of the present invention, which is described below in conjunction with the accompanying drawings and includes the following modules:

Module 1: Weibo feature extraction module. Use two different dimensions of pre-trained BERT encoders to convert Weibo text into word-level and sentence-level features in tensor format; use social information to build a Weibo relationship network, and use LINE to embed Weibo nodes in the Weibo relationship network as relationship features.

Module 2: Weibo feature interaction module. Construct a dual-channel, multi-convolution kernel CNN word-level feature fusion network to interactively fuse word-level features; use relationship features to construct a Weibo similarity matrix, and use the Weibo similarity matrix to guide the interaction of word-level and sentence-level features of Weibo texts; construct an Attention network to perform the first fusion of the interactive word-level features, sentence-level features, and relationship features to obtain a fusion feature.

Module 3: Weibo feature fusion and sentiment classification module. The first fusion feature, the word-level feature after interaction, and the sentence-level feature after interaction are stacked and weighted; an interactive fusion network is constructed to perform secondary fusion of the features after weight assignment; the cross entropy loss function is used as the training loss function, and the back propagation algorithm is used to train the model to obtain the Weibo sentiment analysis model and perform sentiment classification on the Weibo text.

Optionally, module one specifically includes:

(1) Extraction of microblog text and relation features. This includes the extraction of microblog text features and the extraction of microblog relation features.

用户提及

主题标签

)。Suppose all microblogs are represented by C = {c ₁ ,c ₂ ,c ₃ ,...,c _m } (m is the number of microblogs in the dataset). For any microblog c _i , it contains the microblog text s _i and the microblog interaction features x _i (including the user follow relationship

User Mentions

Hashtags

).

(l为句子的长度)作为词级的表示特征。将BERT模型输出的最后一个隐藏层的[CLS]特征

作为句级的表示特征。则经过BERT编码器，输出的文本特征可以表示为：(1.1) Extraction of microblog text features. This includes the extraction of word-level and sentence-level features. In order to obtain local semantic features and global sentence context features in the text, it is planned to extract word-level features and sentence-level features from the text at the same time to obtain vocabulary and context sentiment features. Two pre-trained BERT models with different dimensions are used to complete this task. Specifically, for the microblog text s _i , the features of the state of the last hidden layer in the BERT model are extracted.

(l is the length of the sentence) as the word-level representation feature. The [CLS] feature of the last hidden layer output by the BERT model

As a sentence-level representation feature, after the BERT encoder, the output text feature can be expressed as:

表示第1维度的BERT模型输出的词级特征，

表示第1种维度BERT模型输出的句级特征，Linear(X,y)的意义在于给特征X乘上过度的可训练矩阵，使输出的特征维度映射为y维。最终可以得到所有微博文本的特征表示：Among them, d ₀ and d ₁ are the dimensions of the two BERT pre-training models.

Represents the word-level features output by the BERT model in the first dimension,

It represents the sentence-level features output by the BERT model in the first dimension. The meaning of Linear(X,y) is to multiply feature X by an excessive trainable matrix so that the output feature dimension is mapped to the y dimension. Finally, the feature representation of all Weibo texts can be obtained:

(1.2) Extraction of user interaction features. User interaction feature extraction is divided into two steps: construction of microblog relationship weight network and embedding of microblog relationship weight network.

用户发布微博的提及标签

用户发布微博的主题标签

三个方面构建微博间的关系权重网络。对于依靠主题标签(#)建立的微博关系的规则为：两个微博处于同一主题下，关系权重表示为两个微博共享相同主题的个数。对于依靠提及标签(@)建立的微微博关系的规则为：(1)微博A提及用户B后，用户B发布微博B，则微博A和微博B间有一条权重为1的边；(2)两个微博有相同提及的用户，权重为提及相同用户的个数。对于依靠用户关注建立的微博关系的规则为：(1)用户之间有关注关系且用户发布微博，则这些微博之间有一条权重为1的边；(2)同一用户发布的微博之间有一条权重为1的边。最终可以构建一个微博关系权重网络E(i,j,w)。The construction of Weibo relationship weight network.

The mention tags of the user's microblog

Theme tags for users to post on Weibo

The relationship weight network between microblogs is constructed from three aspects. The rule for the microblog relationship established by the topic tag (#) is: if two microblogs are under the same topic, the relationship weight is expressed as the number of topics shared by the two microblogs. The rule for the microblog relationship established by the mention tag (@) is: (1) after microblog A mentions user B, user B posts microblog B, then there is an edge with a weight of 1 between microblog A and microblog B; (2) if two microblogs have the same mentioned user, the weight is the number of times the same user is mentioned. The rule for the microblog relationship established by user following is: (1) if there is a following relationship between users and users post microblogs, there is an edge with a weight of 1 between these microblogs; (2) there is an edge with a weight of 1 between microblogs posted by the same user. Finally, a microblog relationship weight network E(i,j,w) can be constructed.

Embedding of Weibo relation weighted network. LINE is used to embed nodes in a weighted network as corresponding low-dimensional feature vectors. For the Weibo relation weighted network E(i,j,w) established in the first step, Weibo nodes v _i and v _j , the embedding representation of each Weibo node in the weighted network can be obtained by minimizing the embedding loss function O.

表示微博i的低维嵌入向量，

表示当

视为上下文的向量。生成微博关系嵌入表示为

微博数量为m，微博节点嵌入维度是d_r。其中，没有参与到社交互动的微博节点填充为相同维度的噪声向量

得到m个微博节点的嵌入矩阵表示为

Among them, (i,j,wi _,j ) indicates that there is an edge with weight wi _,j between Weibo i and Weibo j.

represents the low-dimensional embedding vector of Weibo i,

Indicates when

The generated microblog relationship embedding is represented as

The number of microblogs is m, and the embedding dimension of microblog nodes is d _r . Among them, microblog nodes that do not participate in social interactions are filled with noise vectors of the same dimension

The embedding matrix of m microblog nodes is expressed as

Optional, Module 2 specifically includes:

采用CNN提取到的词级特征输出

表示为：(1) Word-level feature fusion network. A dual-channel CNN with multiple convolution kernels of different sizes is used to extract word-level high-order features. Specifically, for the word-level feature vector _extracted from the microblog text si

Output of word-level features extracted by CNN

It is expressed as:

y＝{y ¹ ,y ² ,...,y ^L }

为两个BERT输出到卷积双通道的堆叠，W^l和b^l分别表示第l个卷积核的权重矩阵和偏置，

表示卷积操作；

表示在一个包含h个词的窗口中，第l个卷积核的输出。接着，通过最大池化层找出文本中的重要部分p并在最后一维度连接所有特征。得到最终的词级特征

表示为：Where ^yi represents the output of the i-th convolution kernel,

is a stack of two BERT outputs to the convolution dual channel, W ^l and b ^l represent the weight matrix and bias of the lth convolution kernel respectively,

Represents the convolution operation;

represents the output of the lth convolution kernel in a window containing h words. Next, the important part p in the text is found through the maximum pooling layer and all features are connected in the last dimension. The final word-level feature is obtained

It is expressed as:

p=cat([maxpool(y ¹ ),maxpool(y ² ),...,maxpool(y ^L )])

表示为In order to unify the dimensions, the word-level feature p is mapped to d _r dimensions, that is,

Expressed as

_ps ＝Linear(p, _dr )

也映射为d_r维，即

表示为：Extraction of sentence-level features. The sentence-level features of the first BERT are used as the overall sentence features.

It is also mapped to d _r dimension, that is,

It is expressed as:

(2) Guided interaction of relational features. This includes the construction of microblog similarity matrix, and guided interaction at word level and sentence level.

对于第i个微博节点嵌入

得到微博相似性矩阵M为：(2.1) Construction of microblog similarity matrix. The microblog similarity matrix is constructed using relationship features and is defined as the normalized value of the similarity between two node vectors. The embedding matrix of m microblog nodes is expressed as

For the i-th microblog node embedding

The microblog similarity matrix M is obtained as:

h＝Tanh(sum _r (M _r -I ₍₁₎ *v _noise )*sum _r (M _r -I*v _noise ))

M _corr =ReLU(Tanh((h·h ^T -I)*10))

Where || _vi || _F is the Frobenius norm of vector _vi , · is matrix multiplication, * is element multiplication, and sum _r () represents summing the matrix by row. _{I (1)} is a matrix of all 1s, and I is the identity matrix. M _corr is a correction matrix, which is mainly used to correct the microblog text similarity noise caused by the filling vector v _noise of the randomly filled microblog nodes. Tanh and ReLU are activation functions, as follows:

ReLU(x)=max(0,x)

交互后句级特征

输出分别为：(2.2) Guided interaction at word level and sentence level. For m microblogs, the word-level features after interaction

Post-interaction sentence-level features

The outputs are:

(3) The first fusion of features. The Attention network is constructed to perform the first fusion of the extracted interaction features, word-level features after interaction, and sentence-level features after interaction. Specifically, the microblog relationship feature _Mr is used as the guided query (Q), the sentence-level feature _hcls after interaction is used as the key value (K), and the word-level feature _hs is used as the value (V). These are input into the self-attention mechanism to obtain the first fusion feature of the guided microblog text:

h _att = SelfAtt (Q = M _r , K = h _s , V = h _cls )

Where d _k is the dimension of the key-value vector K. Q, K, V are the corresponding query, key-value, and value vectors.

Optional, Module 3 specifically includes:

(1) Dynamic weighting. The first fusion feature h _att , the word-level feature after interaction h _cls , and the sentence-level feature after interaction h _s are stacked and dynamically weighted. The weighted feature is:

h _stack =Stack([h _s ,h _att ,h _cls ])

h _c =(h _stack ) ^T *[a ₁ ,a ₂ ,a ₃ ]

为堆叠后的特征，[a₁,a₂,a₃]为3个特征的动态加权系数。in,

is the stacked feature, and [a ₁ ,a ₂ ,a ₃ ] is the dynamic weighting coefficient of the three features.

(2) Interactive fusion network. Similar to the word-level feature fusion network, CNN is used to perform a second fusion of weighted features, and the fused feature h is:

h＝{h ¹ ,h ² ,...,h ^L }

Next, the interactive features are obtained through average pooling and concatenated in the last dimension. The fused features are obtained as

h _f =cat([avgpool(h ¹ ),avgpool(h ² ),...,avgpool(h ^L )])

(3) Microblog sentiment classification. Construct a Softmax classifier to complete the text sentiment classification:

Among them, num_class is the sentiment category corresponding to the microblog text.

The back propagation algorithm is used to train the model, and the cross entropy loss function is used as the loss function in the training process. The model is optimized by optimizing the loss function. The expression is:

分别为样本的真实样本概率分布和预测样本概率分布。λ为L2正则化的系数。Among them, J(w,b) is the loss of the whole sample, m is the number of samples, y ⁽ⁱ⁾ ,

are the true sample probability distribution and predicted sample probability distribution of the sample respectively. λ is the coefficient of L2 regularization.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims (7)

1. A social relationship-guided multi-level feature interactive fusion microblog sentiment analysis method, characterized in that the method comprises the following steps: Step 1: Extraction of Weibo text and user interaction features: Preprocessing of Weibo text and user social behavior data, using BERT pre-trained models of different dimensions to extract word-level and sentence-level features of Weibo text, and using LINE to extract relational features of social information, expressed in a tensor format that can be recognized by computers; Step 2: Fusion of word-level features: It is planned to use a CNN with dual channels and multiple convolution kernels of different sizes to fuse word-level features; Step 3: Interaction of word-level, sentence-level, and relational features. It is planned to use the relational features in step 1 to guide the word-level features in step 2 and the sentence-level features in step 1 to interact, and obtain the word-level features and sentence-level features after interaction. Step 4: The first fusion of features: using the relational features of step 1 as the query vector Q, the sentence-level features of step 3 as the key-value vector K, and the word-level features of step 3 as the value vector V as features, construct an attention fusion network to fuse the three features for the first time, and obtain a fused feature; Step 5: Second fusion of features; Dynamic weighting coefficients are used to complete the weight allocation of the first fusion features of step 4, the word-level features after interaction of step 3, and the sentence-level features after interaction; an interactive fusion network is constructed to perform a second fusion of the weighted features; Step 6: Weibo sentiment classification; construct a Softmax sentiment classifier to classify the sentiment of Weibo; use the cross entropy loss function as the training loss function, use the back propagation algorithm to train the model, and obtain the Weibo sentiment analysis model. 2. The social relationship-guided multi-level feature interactive fusion microblog sentiment analysis method according to claim 1 is characterized in that: the specific process of step 1 includes: First, for the microblog text s _i , the features of the state of the last hidden layer in the BERT model are

l is the length of the sentence, which is used as the word-level representation feature; the [CLS] feature of the last hidden layer output by the BERT model

As a sentence-level representation feature; after the BERT encoder, the output text feature is represented as:

Among them, d ₀ and d ₁ are the dimensions of the two BERT pre-training models.

Represents the word-level features output by the BERT model in the first dimension,

It represents the sentence-level features output by the BERT model of the first dimension. The meaning of Linear(X,y) is to multiply the feature X by an excessive trainable matrix so that the output feature dimension is mapped to the y dimension. Finally, the feature representation of all Weibo texts is obtained:

Then, the user interaction features are extracted; the user interaction feature extraction is divided into two steps: the construction of the Weibo relationship weight network and the embedding of the Weibo relationship weight network; the construction of the Weibo relationship weight network; the focus relationship between users is extracted.

The mention tags of the user's microblog

Theme tags for users to post on Weibo

The relationship weight network between microblogs is constructed from three aspects; the rule for the relationship between microblogs established by the topic tag # is: if two microblogs are under the same topic, the relationship weight is expressed as the number of topics shared by the two microblogs; the rule for the relationship between microblogs established by mentioning the tag @ is: (1) After Weibo A mentions user B, user B posts Weibo B, then there is an edge with a weight of 1 between Weibo A and Weibo B; (2) If two microblogs have the same user mentioned, the weight is the number of times the same user is mentioned; The rules for microblog relationships based on user attention are: (1) If there is a following relationship between users and the users post microblogs, there is an edge with a weight of 1 between these microblogs; (2) There is an edge with a weight of 1 between microblogs posted by the same user; finally, a microblog relationship weight network E(i, j, w) is constructed; Finally, embedding of the Weibo relation weight network; using LINE to embed nodes in a weight network as corresponding low-dimensional feature vectors; for the Weibo relation weight network E(i,j,w) established in the first step, Weibo nodes _vi and _vj , the embedding representation of each Weibo node in the weight network is obtained by minimizing the embedding loss function O;

Among them, (i,j,wi _,j ) indicates that there is an edge with weight wi _,j between Weibo i and Weibo j.

represents the low-dimensional embedding vector of Weibo i,

Indicates when

The vector of the context is regarded as; the generated microblog relationship embedding is represented as

The number of microblogs is m, and the embedding dimension of microblog nodes is d _r ; among them, microblog nodes that do not participate in social interactions are filled with noise vectors of the same dimension

The embedding matrix of m microblog nodes is expressed as

3. The social relationship-guided multi-level feature interactive fusion microblog sentiment analysis method according to claim 2 is characterized in that: the specific process of step 2 includes: First, the word-level feature vector _extracted from the microblog text si

Output of word-level features extracted by CNN

It is expressed as: y＝{y ¹ ,y ² ,...,y ^L }

Where ^yi represents the output of the i-th convolution kernel,

is the stack of two BERT outputs to the convolution dual channel, W ^l and b ^l represent the weight matrix and bias of the lth convolution kernel respectively, and ο represents the convolution operation;

Represents the output of the lth convolution kernel in a window containing h words; then, the important part p in the text is found through the maximum pooling layer and all features are connected in the last dimension; the final word-level feature is obtained

It is expressed as: p=cat([maxpool(y ¹ ),maxpool(y ² ),...,maxpool(y ^L )]) Then, to unify the dimensions, the word-level feature p is mapped to d _r dimensions, that is,

Expressed as _ps ＝Linear(p, _dr ) Finally, sentence-level features are extracted; the sentence-level features of the first BERT are used as overall sentence features; that is, sentence features

It is also mapped to d _r dimension, that is,

It is expressed as:

4. The social relationship-guided multi-level feature interactive fusion microblog sentiment analysis method according to claim 3 is characterized in that: the specific process of step 3 includes: The first is the guided interaction of relational features, including the construction of microblog similarity matrix, and guided interaction at word and sentence levels; Then, the microblog similarity matrix is constructed; the microblog similarity matrix is constructed using the relationship features, which is defined as the normalized value of the similarity between two node vectors; the embedding matrix of m microblog nodes is expressed as

For the i-th microblog node embedding

The microblog similarity matrix M is obtained as:

h＝Tanh(sum _r (M _r -I ₍₁₎ *v _noise )*sum _r (M _r -I*v _noise )) M _corr =ReLU(Tanh((h·h ^T -I)*10)) Where || _vi || _F is the Frobenius norm of vector _vi , · is matrix multiplication, * is element multiplication, sum _r () means summing the matrix by row; I ₍₁₎ is an all-1 matrix, I is the identity matrix; _Mcorr is a correction matrix, which is mainly used to correct the microblog text similarity noise caused by the filling vector v _noise of the randomly filled microblog nodes; Tanh and ReLU are activation functions, as follows: ReLU(x)=max(0,x)

Finally, the guided interaction at the word level and sentence level; for m microblogs, the word-level features after interaction

Post-interaction sentence-level features

The outputs are:

5. The social relationship-guided multi-level feature interactive fusion microblog sentiment analysis method according to claim 4 is characterized in that: the specific process of step 4 includes: Specifically, the microblog relationship feature _Mr is used as the guided query (Q), the sentence-level feature _hcls after interaction is used as the key value (K), and the word-level feature _hs is used as the value (V), which is input into the self-attention mechanism to obtain the first fusion feature of the guided microblog text: h _att = SelfAtt (Q = M _r , K = h _s , V = h _cls )

Among them, d _k is the dimension of the key-value vector K; Q, K, V are the corresponding query, key-value, and value vectors. 6. The social relationship-guided multi-level feature interactive fusion microblog sentiment analysis method according to claim 5 is characterized in that: the specific process of step 5 includes: First, dynamic weighting is performed; the first fusion feature h _att , the word-level feature after interaction h _cls , and the sentence-level feature after interaction h _s are stacked and dynamically weighted; the weighted feature is: h _stack =Stack([h _s ,h _att ,h _cls ]) h _c =(h _stack ) ^T *[a ₁ ,a ₂ ,a ₃ ] in,

is the stacked feature, [a ₁ ,a ₂ ,a ₃ ] is the dynamic weighting coefficient of the three features; Next, the interactive fusion network; similar to the word-level feature fusion network, uses CNN to fuse the weighted features for the second time, and the fused feature h is:

h＝{h ¹ ,h ² ,...,h ^L } Finally, the interactive features are obtained by average pooling and concatenated in the last dimension; the fused features are obtained as h _f =cat([avgpool(h ¹ ),avgpool(h ² ),...,avgpool(h ^L )]). 7. The social relationship-guided multi-level feature interactive fusion microblog sentiment analysis method according to claim 6 is characterized in that: the specific process of step 6 includes: First, build a Softmax classifier to complete the classification of text sentiment:

Among them, num_class is the sentiment category corresponding to the microblog text; Finally, the back propagation algorithm is used to train the model, and the cross entropy loss function is used as the loss function in the training process. The model is optimized by optimizing the loss function. The expression is:

Among them, J(w,b) is the loss of the whole sample, m is the number of samples, y ⁽ⁱ⁾ ,

are the true sample probability distribution and predicted sample probability distribution of the samples respectively; λ is the coefficient of L2 regularization.

Images