3.1.1 Supervised learning

There is extensive literature that leverages the combinations of the characteristics described earlier to train a supervised sentiment classification model [49]. In this research area we can roughly distinguish between the state of the art in two main fields (ie, baseline and ensemble models). Concerning the baseline models, several contributions have been provided in the last decade. Among the most recent approaches are those in [50–55].

Dhande and Patnaik [50] proposed a supervised machine learning approach that is able to combine naïve Bayes and neural network classifiers for sentiment categorization. In their investigation, they tried to overcome the attribute independence assumption underlying the naïve Bayes classifier by using a neural network to explicitly represent the dependencies among attributes (ie, words). The resulting naïve Bayes neural classifiers have been shown to be able to achieve promising accuracy with use of a simple unigram representation of user messages.

Chikersal et al. [51] combine two main baseline approaches. The first one is a rule-based classifier that takes a decision about positive, negative, or unknown sentiment, using rules that are dependent on the occurrences of emoticons and opinion words in tweets (ie, lexicons). The second one, based on support vector machines, is trained on semantic, dependency, and sentiment lexicons to identify positive, negative, and neutral messages. The predictions provided by the two baseline approaches are subsequently combined to refine the support vector machine predictions. A further supervised approach based on sentiment lexicon construction was presented in [55]. The main goal is to deal with the problem of unavailability of labeled data by use of SentiWordNet [56]. The proposed framework, SWIMS (for “semisupervised subjective feature weighting and intelligent model selection”), starts by acquiring SentiWordNet as a labeled corpus to extract adjectives, verbs, adverbs, and nouns that are subjective. Then a new feature-weighting mechanism—based on the pointwise mutual information [57]—is used to finally train a supervised support vector machine for sentiment classification.

Even though the traditional supervised learning approaches have addressed several sentiment analysis tasks, most of the recent literature is moving toward a different perspective of the problem. In particular, the current research direction is related to the definition of novel representation spaces, by means of word embeddings [58], for subsequent training of traditional learning models [52–54]. Severyn and Moschitti [52] proposed a new model for initializing the parameter weights of a convolutional neural network [59] based on word embedding to finally train an accurate soft-max sentiment model. The solution they proposed successfully combines two important constituents for sentiment analysis: word embeddings for a rich language model and supervised learning on available data. A similar architecture was presented in [53]. A further investigation that exploits distributed word representations was proposed in [54], where two types of word embeddings were adopted (ie, the skip-gram model [60] and the sentiment-word model [61]) to subsequently train a support vector machine.

Although the above-mentioned approaches represent an important step toward the definition of robust systems, within the sentiment classification research field, none of the classification algorithms consistently performs better than the others and there is no consensus regarding which method should be adopted for a given problem in a given domain.

To overcome this limitation, the ensemble learning paradigm has been investigated recently [62–67]. The idea behind ensemble mechanisms is to take advantage of several independent classifiers by combining them to achieve better performance than the baseline ones.

A first approach to ensemble composition was presented in [62, 63]. The main idea is for one to exploit all the possible baseline models in the hypothesis space by taking into account their marginal prediction capabilities and their reliabilities. One finally derives the optimal ensemble composition by taking advantage of a greedy model selection strategy able to find a good bias-variance trade-off among all the baseline models considered. Lin et al. [64] presented as a novel contribution a criterion for sentiment-based ensemble composition. In their work an approximate algorithm, which is able to exploit accuracy and diversity of baseline classifiers, is designed to tackle the combinatorial problem of classifier selection.

A different approach to ensemble learning is related to the feature space instead of the model space. Wang et al. [65] reported an extensive study of the traditional vector representations for inducing state-of-the-art ensemble methods. In particular, they investigated traditional bagging and boosting approaches [68] over different bag-of-word representations (unigram and bigram) and different weighting schema (Boolean, term frequency, and term frequency–inverse document frequency).

In [66] the hypothesis is that an appropriate feature engineering—known as feature hashing [69]—can lead to more accurate ensemble models. The original bag-of-word space is reduced by the hashing of the features into a lower-dimensional space, allowing multiple features to be mapped to the same hash key. An approach that tries to exploit the potentiality of baseline methods through ensemble learning and enriched word representation was presented by Zhang and He [67], who took advantage of two different feature sets: the first one describes the latent topics of the messages, while the second one is related to word embeddings. These feature sets are then used to tune the contribution of different learners in the ensemble model.

As a general consideration we can highlight that although the approaches based on supervised learning can lead to very high sentiment recognition performance, a drawback is concerned with the necessary human effort to label the data. To overcome this limitation, alternative solutions (grounded on semisupervised and unsupervised learning) have been developed.

3.1.2 Semisupervised learning

Several semisupervised approaches have recently been presented to address the well-known problem related to the acquisition of manually labeled data for sentiment classification. Generally, the semisupervised classification methods can be categorized into two types: approaches based on prior lexical knowledge combined with labeled (and unlabeled) data [23, 70, 71], and bootstrap techniques [72–78].

Concerning the lexical-based approaches, Ju et al. [71] addressed the seed-word selection for semisupervised sentiment classification through a joint lexicon-corpus learning approach. They investigated the (human) costs for annotating words to then measure their informativeness. Both the annotation cost and the informativeness measurement are taken into account to guide the selection strategy for good words for the final sentiment analysis task. Although the pure lexical-based approaches have shown promising results for tackling semisupervised sentiment classification, the most recent literature is focused on sentiment transfer learning. He et al. [23] proposed a transfer learning approach that is able to take advantage of the sentiment knowledge available in a resource-rich (source) language to restore the information lost during the transfer process in the new resource-poor (target) domain. Zhu et al. [70] presented an approach that combines lexicons and labeled and unlabeled data for sentiment transfer across different domains. First, several emotion keywords are used to automatically extract labeled samples and then both the automatically labeled samples from the target domain and the real labeled samples from the source domain are combined to create a new labeled data set. Finally, all the labeled and unlabeled data in the target domain are used to perform cross-domain sentiment classification with a standard label-propagation algorithm.

Regarding the bootstrap techniques, we can further distinguish between self-training [79] and co-training [80]. The main characteristic of self-training approaches is to adapt a predefined polarity lexicon with use of an unlabeled set of social network messages. Among the different self-training solutions for sentiment analysis, the most recent approaches are focused on the enrichment of existing vocabularies with (unsupervised) sentiment lexical items for a subsequent learning phase [72–74, 78].

An additional strategy for semisupervised boosting is represented by the co-training approaches. The key characteristic of such methods is their ability to derive and take advantage of different “views” for the same opinionated text. Liu et al. [77] exploit both textual features (ie, opinionated words weighted according to WordNet-Affect [81]) and nontextual indications (ie, emoticons, temporal indications, and punctuation) to train two supervised classifiers. A further approach was presented in [76], where a semisupervised deep neural network framework was developed to co-train on the feature representation and pattern classification spaces. Yang et al. [75] recently investigated a combination of lexicon-based learning and corpus-based learning in a unified co-training framework.

After analyzing the current state of the art, we can affirm that although semisupervised methods for sentiment classification are still in their infancy, they are becoming ever more popular thanks to their ability to use both (small sets of) a labeled corpus and (huge sets of) unlabeled data.

3.1.3 Unsupervised models

Unsupervised approaches to sentiment classification can solve the problem of domain dependency and reduce the need for annotated training data. Most of the approaches available for unsupervised sentiment classification can be broadly distinguished into lexicon-based methods [82–85] and generative models [86–89].

Concerning the lexicon-based approaches, the first work was presented by Turney [82], who used two arbitrary seed words to compute the semantic orientation of a sentence (measured by pointwise mutual information [57]). An improvement on the above-mentioned approach was by Zagibalov and Carroll [83], who presented a method for the automatic seed word selection. The method requires information only about commonly occurring negations and adverbials to iteratively find sentiment-bearing items.

More recent methods are based on the automatic construction of lexicons. Lu et al. [85] addressed the problem of deriving a sentiment lexicon that is not only domain specific but also aspect dependent. To accomplish this task, an optimization framework was proposed to combine different sources of information for learning context-dependent sentiment lexicons.

Sheng et al. [84] proposed an automatic construction strategy for a domain-specific sentiment lexicon based on constrained label propagation. In particular, a set of candidate sentiment terms is extracted by use of the chunk dependency information and an a priori generic lexicon. Then a set of pairwise contextual and morphological constraints are extracted from a domain corpus and are exploited as prior knowledge to improve the sentiment lexicon construction.

The generative models represent an alternative and more flexible solution to the lexicon-based approaches. The main characteristic of this class of models is that they are able to simultaneously extract aspects and classify sentiments from textual messages. A sentence such as “#iOS7: battery life is a plus, but the security is a big issue!” is an example of different aspects, each having its own polarity, reported on the main target “iOS7.” Several studies that deal with sentiment classification and topic modeling have been proposed in the literature.

A first generative model known as the topic sentiment mixture model to separate topics and sentiment words with use of an extended probabilistic latent semantic analysis model [90] was presented in [87]. Further investigations based on the latent Dirichlet allocation principle [91] can be found in [86, 88], where an aspect and sentiment unification model (ASUM) and a joint sentiment-topic (JST) model were proposed respectively. The JST model is a fully unsupervised model that can capture the topic and sentiment at the same time. The ASUM is a model based on the JST model, characterized by a small adaption of the former one, that introduces the assumption that a sentence in a message can be related to only a specific aspect and sentiment. The main advantage of the JST model with respect to the ASUM comes from its ability to reciprocally reduce the noise of both topic and sentiment generation tasks.

A more recent generative model available in the literature was presented by He et al. [89]. They proposed a dynamic JST model that allows the detection and tracking of topics and sentiment over time: topic and sentiment dynamics are captured by the current time-dependent sentiment-topic word distributions and the word distributions at previous time stamps.

3.2.1 Supervised learning

The first and more popular investigations in the state of the art aimed at combining content and relationships are grounded on the status homophily sociological theory. Assuming status homophily as an underlying principle of online relationships between users implies the assumption that users form ties thanks to demographic factors (ie, race/ethnicity, sex/gender, age, religion, education, occupation, and social class). According to this principle the relationships between users have been modeled in online social networks both as directed and as undirected binary connections: friendships in Facebook,² Google+,³ and Weibo,⁴ and following/follower in Twitter.⁵

Concerning the state-of-the-art approaches that assume relationships based on status homophily jointly with text, we highlight the most recent supervised learning methods [92–94]. The first tentative approach exploiting the statistical relational learning paradigm for polarity classification purposes was presented Rabelo et al. [92], who investigated a relational neighbor classifier to estimate a polarity probability model (based on text ad adjoining friends) and a relaxed collective inference approach to determine the sentiment of the users in the network. Hu et al. [93] presented a mathematical programming formulation that is able to capture the sentiment consistency (by means of user-content connections) and the emotional contagion (by taking advantage of user-user social relations) for networks of reduced size. Recently, an investigation into the emotional contagion in large social networks was presented by Coviello et al. [94], who modeled the emotions of the users as dependent not only on the endogenous and exogenous factors (eg, always being happy and rainfall effect) but also on contagion of groups of friends.

Although the above-mentioned investigations are characterized by consistent performance gains with respect to those approaches based purely on textual content, they strongly assume that the explicitly available relationships (friendships and following/follower connections) unconditionally represent the sentiment agreement between connected users. However, this assumption does not reflect what happens in the real world, where two structurally connected users can have divergent opinions on a given topic. To better capture the sentiment agreement among users, some recent approaches have been grounded on the value homophily theory. Considering this principle as underlying the online relationships between users implies the assumption that users create bonds thanks to attractions and interactions that contribute to their sharing of attitudes, abilities, beliefs, and aspirations. According to this principle the relationships between users have been modeled in online social networks as weighted directed connections: appreciations in Facebook (ie, “like”) and Google+ (“+1”) and retweets in Twitter (“RT”).

Regarding those approaches in the literature that assume relationships based on value homophily together with text, we describe a few recent investigations. Jiang et al. [95] derived a probabilistic Bayesian model that considers the text content posted by a user smoothed by the sentiment labels of neighbors directly connected through a structural connection (ie, a retweet in Twitter). Once the probabilistic model has been trained, a graph-based classification algorithm based on relaxation labeling [96] is used to infer the polarity of unobserved users.

Anjaria and Guddeti [97] proposed a supervised system that exploits the salient features of supervised machine learning algorithms for text-based sentiment classification to subsequently incorporate social network structural features (again a retweet in Twitter) as an approach to sentiment analysis at the user level. Finally, in [98] any kind of social interaction (eg, “like,” comments, and sharing activities) is captured by a sentiment opinion graph. To derive the sentiment orientation of each user of the network, a relaxation labeling process is followed.

3.2.2 Semisupervised learning

Dealing with sentiment classification in networked environments usually requires a fully supervised learning paradigm, where the sentiment orientation must be known a priori to derive suitable predictive models. However, this does not always reflect the real setting of social networks, where some partial information can be grasped to derive a relational semisupervised model and therefore reduce the human effort related to the annotation process. Among the semisupervised techniques able to deal with content and relationships, we can again distinguish between models that include status homophily relationships and approaches that take advantage of value homophily relationships.

Among the models based on status homophily relationships, we highlight the most recent approaches presented in the literature [99–102]. Tan et al. [99] proposed a semisupervised approach to predict the user sentiment by introducing explicitly available undirected user-user relationships (“friendships”) into a text-based factor-graph model. Speriosu et al. [100] proposed enriching the content representation by including directed user-user relationships as features additional to the text ones. The same kind of directed user-user social relationships (eg, “following” and “follower” in Twitter) was exploited in [101] to predict the sentiment orientation of users by a collective inference approach based on a partially labeled network. Tang et al. [102] recently presented a semisupervised version of the supervised model presented in [93]. The main rationale behind this recent approach is (1) content from the same author is likely to be more consistent in terms of polarity than two randomly selected messages and (2) content provided by connected friends is likelier than two randomly selected texts to have the same sentiment.

Concerning the value homophily relationships, only two semisupervised approaches have been found in the literature. Pozzi et al. [103] proposed a semisupervised sentiment learning approach that extends the model presented in [99] by introducing a social network representation based on the concept of an “approval network.” Given a small proportion of users already labeled in terms of polarity, the model predicts the sentiments of the remaining unlabeled users by combining directly in the probabilistic model both the textual information and the retweet-based graph representation.

A subsequent contribution built on the work reported in [103] was presented by Nozza et al. [104]. In their investigation a social network is represented as a heterogeneous graph, where a latent representation of the nodes (both users and posts) is learned to infer the corresponding polarity labels.

3.2.3 Unsupervised learning

Similarly to the semisupervised paradigm, the unsupervised approaches in the literature able to leverage both content and relationships are very preliminary. The work in [105, 106] represents the first tentative approaches to address the polarity detection task assuming relationships based on status homophily. The first explorative analysis in [105] proposed an unsupervised label propagation algorithm for dealing with both explicit and implicit opinion targets. The authors consider posts as nodes in the graph with a corresponding polarity label vector initialized by the hashtags reported in the text. Then at each iteration of the label propagation algorithm, the label vector of one node is propagated to the adjoining ones.

Ou et al. [106] proposed a content and link unsupervised sentiment model as a richer framework able to take advantage of four main components: content, same-user link (ie, two posts are provided by the same user), friend link (ie, the users of two messages are connected by a friendship relationship), and behavior link (two users are connected if any repost, reply, or comment activity is performed). These four ingredients were introduced in a unified probabilistic model, for which parameter estimation and inference can be approached by a maximum likelihood method and Gibbs sampling respectively.

Concerning the value homophily relationships, only one unsupervised approach has been found in the literature. Zhu et al. [107] proposed an unsupervised triclustering framework that is able to analyze both user-level and message-level sentiments through co-clustering of a tripartite graph. The most important contribution is concerned with the finding that making use of mutual dependency among aspects, messages, and user relationship can lead to effective unsupervised sentiment classification models.