Since, this work aims to propose an advanced network embedding framework for implicit sentiment analysis, this section summarizes related works from two aspects: sentiment analysis and network embedding.
Sentiment analysis
Sentiment analysis is one of the most important tasks in the field of natural language processing. The existing text-based studies of sentiment classification can be mainly grouped into two categories: lexicon-based and corpus-based approaches. Lexicon-based approaches mainly use the sentiment polarity associated with the sentiment dictionary to calculate the sentiment polarity of each sentence or document. Corpus-based approaches [
6‐
8] take sentiment classification as a special case of text categorization problem, which utilize machine learning methods to extract reasonable features from texts and feed them into a classifier to predict sentiments. Nowadays, more people are used to express their attitudes on different entities in online social networks, forming user to entity sentiment links. These sentiment links imply positive or negative semantics. Most of current user sentiment analysis literature focuses on making a positive, neutral, or negative sentiment decision according to users’ text descriptions. And more and more attention has been paid to the task of user-entity sentiment analysis. For example, Li et al. [
9] integrate both named entity information and sentiment level information together to form label sequences and adopt an approach based on graphical models. Gan et al. [
10] propose a self-attention based hierarchical dilated convolutional neural network for multi-entity sentiment analysis. Ding et al. [
11] design an entity-level sentiment analysis tool consisting of sentiment classification and entity recognition, which can classify issue comments into < sentiment, entity > tuples. However, these above entity sentiment analysis are mainly rely on text content, and all contain explicit words. Compared with explicit sentiment analysis, the lack of sentiment words in sentences makes the expression of sentiment more euphemistic and these traditional sentiment analysis methods often fail to retrieve users’ hidden real attitudes, making the implicit sentiment analysis more challenging.
Existing works of implicit sentiment analysis mainly focuses on the discrimination of implicit sentiment sentences. To capture implicit structures, recent studies show promising results by learning word embedding with neural language models, such as word2vec [
41] model. The embedded word vector obtained by these methods has better representation and reasoning ability in the semantic space and can be used as the input of various deep learning models, such as LSTM, and CNN models. Kauter et al. [
12] propose a new fine-grained method to identify explicit and implicit sentiment in financial news and reviews. Liao et al. [
5] propose a multi-level semantic fusion method based on representation learning, which can learn and fuse three different levels (i.e. words, sentences and documents) of features to recognize factual implicit sentiment sentences, however, this model is limited by grammatical structure. Bi-directionally long short-term memory [
13] model can effectively capture the semantics of long dependent sequences, and assign different weights to words in sentences automatically via introduction of attention mechanism [
14,
15]. Wei et al. [
16] propose an implicit sentiment classification model based on multi-polarization orthogonal attention mechanism, which can simulate the differences between vocabulary and specific sentiment polarity attention effectively, and improve the performance of implicit sentiment classification. Zuo et al. [
17] propose a context-based heterogeneous graph convolution network model, which first regards the context of the whole document as a heterogeneous graph to maintain its dependency structure, and then employs graph convolution network to obtain the features of implicit sentiment sentences and context.
However, existing implicit sentiment analysis often faces three major challenges: (1) there is a lack of sentiment words; (2) the words are relatively objective and neutral; and (3) the words’ sequences, co-occurrence relationship are neglected. The traditional bag of words-based models, however, can neither accurately capture the meaning and text structure of words in the text, nor effectively represent the semantics of a sentence. In addition, since words are organized into text content by forming phrases, sentences and so on to express users' opinion and the different positions of words in a sentence will have a great impact for a detailed understanding of the text.
Graph-based text representation method is one of the effective methods to solve the above problems. In the graph, nodes represent features and edges represent the relationship between different nodes. Although there are various graph-based text representation models, such as word graphs and ngram-graphs [
18], to better capture the inherent characteristics of text in social media, this paper uses word co-occurrence graph to represent the relationship between words in short text content.
Representing the text in a graph structure that considers all the information for sentiment classification has achieved excellent performance and has attracted considerable attention in recent years. For example, Bijari et al. [
19] propose to represent sentences via graph scheme and Node2vec embedding is used as feature learning algorithms to extract the text features. Zhao et al. [
20] design a bidirectional attention mechanism with position encoding to capture the aspect-specific representations. Gui et al. [
21] propose to build a heterogeneous network by constructing the relationship among words, users and products.
On social networks, users with similar hobbies and concepts often form personalized social communities. For example, Xiao et al. [
22] and Zhao et al. [
23] both consider user structure and model the social relationship of users. In addition, part of the sentiment prediction work also considers user-product-aware information, for example, user product neural network (UPNN) [
24], user product attention (UPA) [
25], user product deep memory network (UPDMN) [
26] and hierarchical user attention and product attention neural network (HUAPA) [
27] have achieved excellent performance in sentiment analysis by considering internal and external connections of users and products.
These above researches demonstrate that both the representation of a text in a graph structure, user social structure and the connections between users and entities are effective for sentiment analysis and opinion mining. These studies inspire this work to take into account both aspects in the proposed framework.
Network embedding
The purpose of network embedding is to represent high-dimensional and sparse vector space with low dimensional and dense vector space. The learned features can be used in machine learning tasks such as classification, regression, clustering and so on. Network embedding methods mainly include matrix eigenvectors calculation-based methods and neural network-based methods. Locally linear embedding (LLE) [
28] and Laplacian Eigenmaps (LE) [
29] are conventional matrix eigenvectors calculation-based methods. Nowadays, neural network-based embedding methods, such as DeepWalk [
30], which inspired by the neural network language model Word2vec [
31] achieves excellent performance and has been widely used. LINE [
32] proposes to establish the first- and second-order proximity between nodes, and can be used in large networks. Node2vec [
33] designs a biased random walk procedure to learn a mapping of nodes that maximizes the likelihood of preserving network neighborhoods of nodes. SDNE [
34] uses auto-encoder to capture network structures and learn user representations.
The above researches are network embedding models and methods. In addition, there are also some literatures focusing on embedding representation algorithms for different entities. For example, Zhang et al. [
35] propose a new optimization model to capture the hidden relationship between item content features for cold-start and content-based recommendations. He et al. [
36] propose the bipartite graph neural network (BGNN), which consists of two central operations as interdomain message passing and intra-domain alignment to learn the embedding representations of user and item nodes. Wang et al. [
37] propose SHINE model, which utilizes deep neural network structure to learn the low-dimensional representation of nodes to predict sentiment links between users and celebrities. Yuan et al. [
38] propose to fuse multi-networks information to learn the nodes’ low-dimensional embedding representation for user behavior classification.
The above studies on network embedding mainly focus on the field of recommender systems and user behavior analysis. The proposed scheme is different from these studies in two aspects. First, this work utilizes network embedding to address a different problem. In particular, this paper models the implicit sentiment analysis problem from special link prediction angle and explore user-entity modeling to find the implicit sentiment in social network comments. Second, this different problem raises new challenges that are not addressed in existing studies. In particular, implicit sentiment analysis requires the proposed embedding approach to be able to capture rich information from different networks, such as user social network, user-entity sentiment network, and word graph-based text level comments. These networks are highly heterogeneous. To address these challenges, a novel multiplex network embedding method called MEISP is proposed in this work. Specifically, deep neural network architecture is introduced for embedding these heterogeneous networks, and a multilayer perceptron structure-based early fusion method is adopted to fuse the text information and social network information. Furthermore, the proposed scheme can also extract nodes’ highly nonlinear representations while preserving the original networks structure.