Deep learning for aspect-based sentiment analysis: a review

Deep learning (DL)

Deep learning can learn the intrinsic connections and potential patterns among large amounts of data, where deep neural networks (DNN) is the fundamental neural network architecture and many variants of DNN arise.

Deep neural networks (DNN)

Figure 3 represents the architecture of the neural network, which is composed of an input layer, a hidden layer and an output layer. In the hidden layer, data from the input layer is added linearly and transformed nonlinearly with activation functions (e.g., Sigmoid, ReLu, tanh). The hidden layer output performs similar operations at the output layer. The aforementioned process is called forward propagation, while the opposite process is called backward propagation. In the backward propagation process, the model adopts the gradient descent method and updates the parameters by minimizing the loss function, which can be used for subsequent prediction. As shown in Fig. 4, the difference between NN and DNN is that DNN has more hidden layers (at least two layers theoretically). The main idea of DNN is to regard the output of the previously hidden layer as the input of the current hidden layer to obtain more abstract high-level features.

In order to avoid gradient disappearance or gradient explosion, weight initialization in neural network is also an important task. Existing initialization approaches mainly include uniform distribution initialization, Gaussian distribution initialization, Xavier initialization (Glorot & Bengio, 2010), Kaiming initialization (He et al., 2015). Liu, Liu & Zhang (2022) proposed that an improved and extended weight initialization method with asymmetric activation function, which can expand the selection range of the activation function and improve the performance of the network.

Aspect-based sentiment analysis (ABSA)

Attention mechanism

Attention mechanism means that people will pay different attention to different aspects of the opinion target, which is widely used in many fields (e.g., computer vision (CV), natural language processing (NLP)).

The main idea of the attention mechanism is to find out the crucial information of the task. Specifically, given a query Q and a series of key-value pairs {(K₁, V₁), …, (K_n, V_n)}, attention mechanism aims to calculate the similarity between the query and each key, then sum up with a corresponding value to obtain the corresponding attention weight. The relevant formula is as follows:

(5) $$Attention(Q,K,V)=\sum _{i=1}^{L}Similarity(Q,K_i)\ast V_i$$

Word embedding

Word embedding is a vector of real numbers in a predefined vector space, which can represent word semantics. There are mainly two representation methods: one-hot encoding and distributed representation. One-hot encoding represents a word as a vector, where the dimension is the size of the vocabulary, the word’s corresponding position is set to 1, and others are all set to 0. Due to the definition of one-hot encoding, vocabulary words are independent of each other, so one-hot encoding can’t express the semantic relevance between words. Besides, the vector’s dimension depends on the vocabulary size, large vocabulary may cause sparse matrix and dimension disaster. Therefore, distributed representation (e.g., Word2Vec (Mikolov et al., 2013), GloVe (Pennington, Socher & Manning, 2014), ELMo (Peters et al., 2018), BERT (Devlin et al., 2019)) is proposed to address these problems, which is low-dimensionality and can express the semantic correlation between words.

In addition, some improvements toward word embedding help ABSA tasks to obtain better performance. Akhtar et al. (2018) minimized the influence of data sparsity by using bilingual word embedding (obtained by learning a bilingual parallel corpus), that is, replacing the out of vocabulary (OOV) word with the nearest translation form in the shared vector space. Liang et al. (2019a) used the sparse coefficient vector to select highly relevant words, based on which the embedding representation of the target and aspect can be adjusted.

Aspect term extraction (ATE)

Aspect term extraction (ATE) can be defined as follows: Given a sentence with n words S = (x₁, x₂, …, x_n) and we denote L_a = {B-AT, I-AT, O-AT} as the output label. For each word x_i, the task aims to predict its corresponding label a_i ∈ L_a. Most previous researches were implemented based on topic models (Shams & Baraani-Dastjerdi, 2017) and conditional random field (CRF) (Jakob & Gurevych, 2010; Shu, Xu & Liu, 2017). However, these methods have their limitations: CRF requires lots of features to work well and topic models usually do not produce highly consistent aspect terms. Therefore, deep learning methods have been proposed to address these shortcomings.

ATE task can be classified into two subtasks: opinion target extraction (OE) and aspect category detection (ACD). For example, for the sentence “The food is so good and so popular that waiting can really be a nightmare.”, the OE task extracts aspect terms (“food” and “waiting”), and the ACD task extracts aspect category (“food” and “service”). Many studies have researched the ATE task (Table 3).

Most of the existing research tends to study the OE task. Poria, Cambria & Gelbukh (2016) proposed a non-linear, supervised CNN with linguistic patterns to process OE task. Tran, Hoang & Huynh (2019) proposed a model that combined BiGRU and CRF, where BiGRU considers the long-distance dependency relationship of sentences and CRF considers the relationship of label transfer. Xu et al. (2018) employed two types of pre-training embedding (i.e., general embedding and domain-specific embedding) on CNN to extract aspects. Wu et al. (2018) proposed a hybrid unsupervised method to solve OE task. Specifically, the model first uses chunk-level linguistic rules and a domain-related filter to extract candidate opinion targets, then uses the text with extracted chunks as pseudo-labeled data to train the GRU network. Ma et al. (2019) formalized ATE task as a sequence-to-sequence (Seq2Seq) learning task, gated unit network, and position-aware attention mechanism are also designed to capture semantics and dependence relationship between aspects and context. Based on the theory that the appearance patterns of aspects in the global context (sentence level) and local context (between the adjacent words) are usually different from other words, Liao et al. (2019b) coupled the global and local representations to find aspects.

Since the OE task and the ACD task are highly interrelated, some studies leverage multi-task learning methods to deal with these two tasks. Xue et al. (2017) proposed a multi-task learning model based on RNN and CNN to solve them at the same time. Wu et al. (2019) proposed a multi-task learning model based on QA-style reviews with character-level embedding and attention mechanism.

Opinion term extraction (OTE)

OTE task extracts opinion terms from corpus, which are rarely studied (Table 4). Previous works are usually implemented based on a dictionary. For example, for the sentence “The food is so good and so popular that waiting can really be a nightmare.”, opinion terms (“good”, “popular”, “nightmare”) can be found by looking up the dictionary since it contains sentiment (positive, positive, negative).

Since aspect terms and opinion terms exist one-to-many correspondence, Fan et al. (2019) proposed a novel sequence labeling subtask named Target-oriented Opinion Words Extraction (TOWE). Formally, given a sentence with n words S = (x₁, x₂, …, x_n) and an aspect a, we denote L_o = {B-OT, I-OT, O-OT} as output label. For each word x_i, the task aims to predict its corresponding label o_i ∈ L_o. Fan et al. (2019) proposed IOG model, which combined global context and the output of IO-LSTM to find the corresponding opinion terms, where IO-LSTM transmits aspect information and context information mutually.

5.3 Aspect-opinion term co-extraction (AOTCE)

In the broad sense, the AOTCE task extracts both aspect and opinion terms at the same time. With the development of demand, this task is transferred from the AOTCE task to the aspect-opinion pair extraction (AOPE) task (Table 5).

AOTCE task processes two sequence labeling tasks for identifying both aspect terms and opinion terms separately. For example, for the sentence “The food is so good and so popular that waiting can really be a nightmare.”, the task can find aspect terms (“food” and “waiting”) and opinion terms (“good”, “popular”, “nightmare”) at the same time. Most previous works utilized the interactive information between aspect terms and opinion terms. Wang et al. (2016a) obtained high-level features with dependency-tree based on RecNN and CRF, where information double propagated between aspect and opinion terms. Wang et al. (2017) proposed a novel model, named coupled multi-layer attentions (CMLA), whose main idea is to capture the complex relations (direct and indirect) between aspect and opinion terms with tensor operations. On the basis of CLMA, Yu, Jiang & Xia (2018) designed a global inference module to capture the syntactic relations between aspect and opinion terms explicitly, which introduces multiple syntactic constraints (i.e., intra-task constraints, inter-task constraints and lexicon constraints).

However, opinion terms may match the wrong aspect for the AOTCE task, which can further lead to the wrong judgment of sentiment classification. Thus, on the basis of the AOTCE task, the AOPE task is proposed to solve the problem above by pairing aspect and opinion terms. For example, we further obtain the aspect-opinion pair (“food”-“good”, “food”-“popular”, “waiting”-“nightmare”) from the example above. Zhao et al. (2020) proposed a novel model SpanMlt, which processes the sentence at span-level and contains term and relation scorer to identify aspect terms, opinion terms and its possible relations. Chen et al. (2020) designed the double-channel network to extract opinion entities and relations simultaneously, two synchronization mechanisms is proposed to make full use of extracted information.

Summary

In this section, we discuss the subtasks of IE task—ATE, OTE and AOTCE (Fig. 7).

ATE extracts aspect terms from the corpus. ATE task is classified into OE and ACD according to whether the aspect terms or the aspect terms category is extracted. Existing literature usually regard ATE task as sequence labeling problem, using deep learning models to label each word as aspect words or non-aspect words. In addition, models can achieve better performance by enriching word embedding, considering highly correlation between OE and ACD tasks.

OTE extracts opinion terms from the corpus, which is a rarely studied task. To be simple, the TOWE task has been designed to find one-to-many correspondences between the given aspect and opinion terms.

We also discuss the evolution of the AOTCE task, which transferred from joint extraction to pair extraction. Most works about the AOTCE task propagated the interactive information between ATE and OTE tasks to obtain higher-level features, but models may mismatch aspect terms and opinion terms. So AOPE task has been designed to match aspect terms and opinion terms correctly based on the results of the AOTCE task. Table 5 shows the application of DL in the AOTCE task. We can find that the result of the AOPE task is better than AOTCE in general, its main reason is that the interaction information between ATE and OTE is considered.

Aspect-context relation sentiment classification

Aspect-context relation sentiment classification mainly captures the interaction information (e.g., semantic information, position information, syntax information) between aspect and context, which is used to learn aspect-specific feature representations or aspect-specific sentence representations. Table 6 shows the related work.

Every existing work almost takes semantic information between aspect and context into consideration. Many studies use attention mechanism to integrate aspect information into context. Wang et al. (2016b) proposed attention-based LSTM with aspect embedding (ATAE-LSTM) based on attention-based LSTM (AT-LSTM), where aspect information is integrated into the input and the process of attention weight calculation. Based on ATAE-LSTM, Tay, Tuan & Hui (2018) proposed aspect fusion LSTM (AF-LSTM), which adopted cyclic convolution and cyclic correlation to model the aspect-context relationship. Ma et al. (2017) proposed interactive attention networks (IAN) to interactively learn attention weights of context and aspect to generate representations of aspect and context respectively. Based on coarse-grained attention mechanisms proposed by IAN, Fan, Feng & Zhao (2018) combined it with fine-grained attention mechanisms to better model word-level interaction between aspect and context. To tackle weak interaction in attention mechanism, Liu & Shen (2020) designed a multi-element attention mechanism to capture precise semantic relation among aspect, context, and hidden state. However, the attention mechanism pays much attention to frequent opinion terms and ignores the infrequent ones, which may introduce noise. Xue & Li (2018), Huang & Carley (2018), and Shuang et al. (2020) used filters and gate mechanism to alleviate the problem by controlling the spread of sentiment information in context. Tay, Tuan & Hui (2017) integrated parameterized neural tensor synthesis or holographic synthesis into the memory selection operation to capture binary interactions between aspect and context. Different from using aspect independent (weakly related) encoders to generate sentence representations, Liang et al. (2019b) improved feature selection and extraction capabilities under the guidance of aspects.

Some studies consider location information to achieve better performance. Target-dependent LSTM (TD-LSTM) (Tang et al., 2016a) split the text according to the position of aspect, both of which is end with aspect and process with LSTM. Target-connection LSTM (TC-LSTM) improved TD-LSTM by concatenating the aspect vector with the context vector. Similarly, Zhang et al. (2019) proposed an efficiency model, which converges at least four times faster and uses co-attention matrices to capture precise semantic relation among aspect, left context, and right context. We can observe that the closer to the aspect, the more correlation information words may contain, so location information can be obtained by assigning attention weight to different contexts according to their distance from the aspect (Tang, Qin & Liu, 2016b; Chen et al., 2017b; Li et al., 2018; Fan, Feng & Zhao, 2018; Shuang et al., 2019; Yin et al., 2019). Besides, the capsule network contains location information, which encodes the spatial relationship between features. Jiang et al. (2019) proposed CapsNet and CapsNet-BERT to encode spatial information as features. Du et al. (2019) introduced the interactive attention mechanism in the capsule routing process to model the semantic relation between aspect and context.

Some studies also consider syntactic information to achieve better performance. Current studies transformed sentences into dependency graph or dependency tree structure, then applied GNN to this structure. Instead of traditional bottom-up dependencies, Zhang & Zhang (2019) used three information exchange channels (i.e., self-to-self, top-to-down, bottom-to-up) to obtain more consistent phrase-level sentiment prediction. Huang et al. (2020), Zhang, Li & Song (2019), and Sun et al. (2019) built GCN over the dependency tree to propagate sentiment features or dependency tree from important syntax neighborhood words to aspect and further exploit syntax information and long-distance word dependencies.

Hybrid methods

In addition to aspect-context relation, existing studies improve the performance of the model by mixing other different relations (e.g. aspect-aspect relation, relations contained in language structure) and introducing external knowledge and tools. Table 7 shows the related work.

Aspect-aspect relation usually appears in multi-aspect sentences, where aspects arranged in order will show a high degree of sentiment relevance and interaction. So it can provide sufficient sentiment classification information. Hazarika et al. (2018) generated aspect-specific sentence representations for each aspect and modeled dependency relationship between aspects according to the occurrence order. Based on Hazarika et al. (2018)’s work, Majumder et al. (2018) proposed a memory network with a multi-hop attention mechanism to obtain more accurate representation by repeatedly matching aspect-specific sentence representation with others. Besides, Fan, Feng & Zhao (2018) designed an aspect alignment loss to enhance attention weights difference among aspects.

Relations contained in language structure contain much important information, which can further help sentiment classification task. Ruder, Ghaffari & Breslin (2016) proposed a hierarchical LSTM (H-LSTM), which stacked a review-level Bi-LSTM on top of sentence-level Bi-LSTMs to model intra- and inter-sentence relations simultaneously. According to the hypothesis that aspect target sequence can control the final sentiment classification results, Peng et al. (2018a) learned Chinese aspects from three different levels of granularity: radical, character, and word.

External knowledge and tools can provide additional prior knowledge and help model focuses more on the important information. Kumar, Kawahara & Kurohashi (2018) proposed a two-layered attention network similar to H-LSTM, which generates better embedding with WordNet. Chen & Huang (2019) generated the knowledge vector with the sentimental knowledge graph. Due to the lack of aspect-level datasets and the easy access to document-level datasets, Chen & Qian (2019) proposed the transfer capsule network (TransCap), which transfers document-level knowledge to aspect-based sentiment classification tasks. Hu et al. (2019b) designed orthogonal regularization and sparse regularization and Bao, Lambert & Badia (2019) used lexicon information and attention regularizer to restrict the allocation of attention weights and provide key information. Besides, some studies (Jiang et al., 2019; Huang et al., 2020) integrated pre-trained BERT into the model to further improve the performance.

Summary

In this section, we discuss the SC task according to the relationship between different targets (e.g., aspect, context, language structure). Aspect-context relation is the most fundamental relation for sentiment classification, it learns aspect-specific feature representations or aspect-specific sentence representations with the help of the interaction information (e.g., semantic information, position information, syntax information) between aspect and context. Aspect-aspect relation, language structure relations, and external knowledge and tools provide extra information, to further improve the performance of the model. Tables 6 and 7 shows the performance of proposed model above. We can find that binary classification accuracy is higher than ternary classification accuracy, because sometimes it’s difficult for neural networks to distinguish neutral sentiment. Besides, relations between different targets can help us improve the performance.

Multi-task learning

The main idea of the multi-task ABSA is to capture the complex interactive relations between its important subtasks of aspect-based sentiment analysis and obtain enriched feature representations.

Multi-task ABSA can be divided into aspect sentiment pair extraction (ASPE), aspect sentiment triple extraction (ASTE), aspect sentiment quadruple extraction (ASQE). Specifically, ASPE task aims to extract opinion pair (aspect term, sentiment polarity), ASTE task aims to extract opinion pair (aspect term, opinion term, sentiment polarity), ASQE task aims to extract opinion pair (aspect term, aspect category, opinion term, sentiment polarity). In the early approaches, ATE task and SC task were considered important subtasks of multi-task ABSA, also known as the ASPE task. Luo et al. (2019) designed a cross-sharing unit (CSU) to take full advantage of the interaction between ATE and SC tasks. With the increasing demand for sentiment analysis, multi-task ABSA is no longer limited to the two important subtasks of ATE task and SC task. OTE task is also an important subtask of multi-task ABSA because aspect terms express their sentiment through opinion terms, resulting in the ASTE task. Peng et al. (2020) first proposed the ASTE task, which transformed ATE, OTE, and SC tasks into a unified task and provided a complete one-time solution. Chen & Qian (2020) proposed relation-aware collaborative learning (RACL), a multi-layer multi-task learning framework with a relation propagation mechanism, to make full use of relations between ATE, OTE, and SC tasks. Zhang et al. (2020) proposed a multi-task learning framework to jointly extract aspect terms and opinion terms and simultaneously parses sentiment dependencies and used biaffine scorer (Dozat & Manning, 2017) to capture the interaction of two words in each word pair. Chen et al. (2021b) proposed an approach by considering the rich syntactic dependence and semantic word similarity in sentences (such as self-interaction relations, neighbor relations, and dependency relations), and adopted GraphSAGE (Hamilton, Ying & Leskovec, 2017) to obtain the rich feature representation of words. Since there are too many aspect terms in the corpus, and the same aspect term may have different descriptions, Cai, Xia & Yu (2021) proposed ASQE task, took ACD task as an important sub-task of multi-task ABSA, and proposed four baseline methods.

Tagging scheme

Previous studies regarded multi-task ABSA as the sequence tagging problem, which can be classified into joint mode and collapsed mode. Specifically, joint mode treats N important sub-tasks as N independent sequence tagging problems, while collapsed mode treats them as one independent sequence tagging problem. Table 8 shows the difference between joint mode and collapsed mode. Given the example sentence, assume that ATE task uses the BIO tagging scheme BIO = {B − Begin, I − Inside, O − Outside} and the SC task use L = {POS, NEG, NEU } to label each word, joint mode will label the word “food” “B” for ATE task and “POS” for the SC task, while collapsed mode will label “B-POS”.

Since the multi-task ABSA can be decomposed into many different types of sub-tasks (e.g., extraction task, pair task, and classification task), how to formalize the multitask ABSA task into a unified tagging scheme is another great challenge. Some novel tagging schemes are proposed to solve the problem. In order to capture the many-to-one and one-to-many relationships between aspect terms and opinion terms, Wu et al. (2020) proposed a novel tagging scheme, grid tagging scheme (GTS), to label the relations of all word pairs with tags C = {N − None, A − Aspect, O − Opinion, Pos − Positive, Neg − Negative, Neu − Neutral}. Xu et al. (2020) proposed a novel tagging scheme, position-aware tagging scheme (JET), which take sentiment polarities, opinion terms and position information into consideration.

In addition, current work also translates the multi-tasking ABSA task into other forms of problems. Yan et al. (2021) transformed ASTE task into sequence to sequence (Seq2Seq) problem and solved it using BART (Lewis et al., 2020), which is a pre-trained Seq2Seq model. Mao et al. (2021) converted ASTE task into two machine reading comprehension (MRC) problems, which were solved by joint training of two Bert-MRC models with shared parameters. Chen et al. (2021a) transformed ASTE task into a multi-turn MRC (MTMRC) problem and solved it with the proposed Bidirectional machine reading comprehension (BMRC) framework. Zhang et al. (2021) regarded the ASQP task as a paraphrase generation process, using the formula “Category is Sentiment because Aspect is Opinion.” to restatement statement.

Summary

In this section, we briefly discuss multi-task ABSA and its main challenges. Different from pipeline manner ABSA, multi-task ABSA improved the model’s generalization ability by making full use of the independent and joint information of its subtask task, which can be classified into ASPE task, ASTE task and ASQE task. Since multi-tasking ABSA encompasses multiple types of tasks, performing it within a unified framework is also a challenge. Table 9 shows the current research in multi-task ABSA.

Implicit sentiment analysis

There are 20% to 30% of sentimental expressions have no explicit sentimental words, so implicit sentiment analysis, which studies the sentiment of texts without sentimental words, has gradually become the focus of many scholars. Implicit sentiment is usually expressed by fact-implied and linguistic rhetoric, where 72% of implicit sentiment sentences are fact-implied ones. Specifically, fact-implied expresses the sentiment by describing the facts, while linguistic rhetoric expresses the sentiment through linguistic rhetoric, such as metaphors and rhetorical (Liao, Wang & Li, 2019a). For example, “It’s only five a.m. and music is already playing in the park.” is a fact-implied, which expresses people’s complaints about noise, because it is when people are sleeping; “He is a snake.” is a rhetorical expression, which expresses his insidious and cunning characteristics.

There have been some works on the implicit sentiment analysis task, its main challenge is how to perceive sentiment and how to integrate external knowledge into the model elegantly. Liao, Wang & Li (2019a) analyzed fact-implied implicit sentiment and found it is usually affected by sentiment target, context semantic background and sentence structure, then a multi-level semantic fusion method is proposed to learn the above features. Wei et al. (2020) proposed a BiLSTM model with multi-polarity orthogonal attention to capture the differences between words and sentiment polarity, orthogonal restriction is also designed to maintain the difference during optimization. Su, Tian & Chen (2016) describes an algorithm to interpret the Chinese nominal and verbal metaphors based on latent semantic similarity, where latent semantic similarity is based on WordNet. Peng, Su & Chen, 2018b) proposed target attention-based LSTM (TRAT-LSTM), which introduced the interaction between aspect and context to obtain reliable sentiment-related classification features. Based on TRAT-LSTM, Su et al. (2019) constructed a cultural-related aspect lexicon and proposed culture-related attention-based LSTM (CRAB-LSTM), which also considered the knowledge of Chinese culture. Mao, Lin & Guerin (2019) using linguistic theories (MIP and SPV) to directly inform the design of DNN for end-to-end sequential metaphor identification. Bouazizi & Ohtsuki (2016) proposed a pattern-based approach and four sets of features to detect sarcasm on Twitter.

Domain adaptation

Sentiment classification is widely known as a domain-dependent problem. In order to learn and obtain an accurate domain-specific sentiment classifier, a large number of labeled samples are needed, which are expensive and time-consuming to annotate (Yuan et al., 2018). Moreover, different datasets in different fields are often in great differences, which makes the effects applied to different fields uneven. Domain adaption methods transform learned knowledge or patterns from a certain domain to a new domain, which requires less supervised training data.

Mining commonalities and differences between domains is a typical method to complete the task. With the help of domain knowledge, Yuan et al. (2018) modeled the task at the feature-level and selected the most domain-related sentimentally discerning features in the sharing the feature extraction layer. Although domain-specific information is not transferable, learning domain-specific representation can help to learn domain-invariant representations, Hu et al. (2019a) used an orthogonal domain-related task (e.g., ATE task) to obtain domain-specific information, so as to further extract domain-invariant features.

Same expressions towards different aspects in different domains may produce different sentiments, so it’s necessary to bridge the gap between domains and find more discriminative features. Wang & Pan (2018) effectively reduced word-level domain offsets with syntactic relations, which are also used as cross-domain invariant “axis information” to establish structural correspondences, then an auxiliary task is generated to predict the syntactic relations between any two adjacent words in the dependence tree. Hu et al. (2019a) designed a domain classifier that is executed in a typical adversarial way, where the classifier tries to determine which domain the domain-invariant features come from, while other modules try to deceive the classifier. Qu et al. (2019) implemented category-level alignment on the premise of global marginal alignment to generate more discriminative features. Li et al. (2019b) proposed selective adversarial learning (SAL) method to perform local semantic alignment of fine-grained domain adaptation, so as to obtain a better sharing feature space.

Most current works regarded the ABSA task as a domain-dependent problem. Existing domain adaption methods can only solve the problem of data scarcity in some specific domains, so it’s interesting to improve the model’s generalization ability and make it adapt to multiple domains simultaneously.

Multimodal sentiment analysis

Multimodal sentiment analysis aims to identify sentiment from multiple modals (e.g., text, image, video), its key challenge is how to fusion the data from different modals. There are mainly three combination strategies: early fusion, intermediate fusion and late fusion. Specifically, early fusion directly integrates multiple sources of data into a single feature vector, late-fusion aggregates the decisions from multiple sentiment classifiers which are trained by only single modal data, intermediate fusion’s fusion process is conducted in the intermediate layers of the neural networks (Huang et al., 2019).

Most studies adopt Intermediate fusion to integrate the information. Huang et al. (2019) integrated intermediate fusion and late fusion into a unified framework to exploit the discriminative features in texts and images, then explore the internal correlations between different modalities. Truong & Lauw (2019) proposed a visual aspect attention network (VistaNet), which uses images as the attention and uses visual information as the alignment of sentence level. Chen et al. (2017a) performed modality fusion at word-level and proposed gated multimodal embedding LSTM with temporal attention model (GME-LSTM(A)), which used gating mechanism trained using reinforcement learning. Xu, Mao & Chen (2019) proposed multi-interactive memory network (MIMN), where two interactive memory networks capture the multiple interactive influences between textual and visual information.

An idea may come up naturally that we can improve the model’s performance by integrating more modalities, but it may introduce noise when modalities convey the opposite sentiment. Building some practicable principles helps us identify which modal data is useful may be a good idea to handle these complex scenarios.

Twitter sentiment analysis

Informal texts don’t have a unified structure and many omissions (e.g., subject, verb, object) and ambiguities (e.g., perhaps, perhaps) may appear, therefore, working with informal text on social media(e.g., Twitter) is a new challenge. The SemEval series dataset also provides specialized datasets for the Twitter sentiment analysis task (e.g., SemEval-2014 task 9 (Rosenthal et al., 2014), SemEval-2015 task 10 (Rosenthal et al., 2015), SemEval-2016 task 4 (Nakov et al., 2016) and SemEval-2017 task 4 (Rosenthal, Farra & Nakov, 2017)).

There are also some works on Twitter sentiment analysis. Vosoughi, Zhou & Roy (2015) analyzed the variation of tweet sentiments across time, authors, locations and used a Bayesian approach to incorporate the relationship between these factors and tweet sentiments into standard n-gram based Twitter sentiment classification. Caragea, Dinu & Dumitru (2018) explored a range of deep learning models at both Twitter and user level and studied the verb tense usage as well as the presence of polarity words in optimistic and pessimistic Twitter.

Multilingual application

With the popularity of social platforms, the focus of sentiment analysis has shifted from English to multilingualism. Peng et al. (2018a) studied Chinese aspect from three different levels of granularity: radicals, characters and words. Dashtipour et al. (2020) proposed a concept-level hybrid framework with linguistic rules and deep learning methods for Persian sentiment analysis. Elshakankery & Ahmed (2019) proposed a hybrid incremental learning method (HILATSA) based on lexicon and machine learning for Arabic Twitter sentiment analysis. Pasupa & Seneewong Na Ayutthaya (2019) compared the performance of several traditional deep learning models on Thai children’s stories. In summary, the challenge of multilingualism is to find the features of specific language (e.g., semantic and syntactic).

Future expectations

According to the above discussion, we can find that the research direction of aspect-based sentiment analysis mainly includes two aspects: on the one hand, the research of basic methods; on the other hand, the improvement of task generalization ability under different task scenarios.

For the study of ABSA task basic methods, knowledge accumulation becomes very important with the continuous mining of rich semantic information in the text. The knowledge graph is conducive to language understanding and can be widely paid attention to in the future.

In order to improve the generalization ability of tasks in different task scenarios, the existing methods are studied from the perspectives of the platform (e.g., e-commerce platform, social media platform), language (e.g., Chinese, English), and expression (explicit and implicit). Future research is likely to focus on increasingly complex real-world scenarios.