Deep learning-based techniques have achieved significant improvement on TER tasks [11]. Word embedding techniques are aimed at learning latent, low-dimensional representations from the language structure and alleviating the problems of sparse features and high-dimensional representation in traditional bag-of-words models. Some well-established embedding models are wildly used in many NLP tasks, such as Word2vec [12, 13] and GloVe [14]. They are trained on a large scale of unlabeled data and aimed to capture fine-grained syntactic and semantic regularities. Recently, the emergence of pretrained language model opened the pretraining era in the NLP field. The pretrained language models provide useful general knowledge and can be fine-tuned to almost all downstream tasks. CoVe and ELMo [15] generate dynamic and context-sensitive word embedding, by which the same words with different contexts are given different word vectors. They greatly alleviate the occurrence of ambiguity. BERT utilizes a large amount of unlabeled data during the training, which helps the model to learn useful linguistic knowledge. Bert performs well in encoding contextual grammatical knowledge and has achieved satisfactory results in many NLP task.

Emotion label correlation is always a critical problem in TER tasks. Deep canonical correlation analysis (DCCA) performs well in feature-aware label embedding and label-correlation aware prediction [16, 17]. Some studies try to explore the label correlations by transforming multilabel classification tasks into ranking or generation tasks. In [18], they transform multilabel emotion classification tasks into a relevant emotion ranking task, which is aimed at generating a ranked list of relevant emotions according to emotional intensity. In [19], multilabel classification is regarded as a sequence generation task, and Sequence Generation Model (SGM) is proposed to learn label correlations. Some approaches attempt to estimate label correlations by the modification of loss function implicitly. The label-correlation sensitive loss function is first proposed in [20] with the BP-MLL algorithm. Joint binary cross-entropy (JBCE) loss [21] is proposed to train the joint binary neural network (JBNN) to capture label relations. To reduce the computational complexity, partial label dependence can also contribute to this task, which is demonstrated in [22]. A semisupervised multilabel method is proposed in [23], while label correlations are incorporated by modifying the loss function. Multilabel classification and label correlations learning can also be realized in a joint learning framework [24, 25].

Emotion recognition of each sentence in hierarchical texts, such as document or dialogue, highly depends on contextual cues, and context modeling is indispensable. To realize contextual emotion recognition, not only need to model the current sentence but also the contextual sentences, which helps to know the overall emotional tendency of the document or dialogue.

Context modeling architectures can mainly be summarized into two kinds: flatten context modeling and hierarchical context modeling. By flatten context modeling, context sentences and current sentence are concatenated, and all tokens are flattened into a word sequence. This sequence is fed into neural networks for contextual information extraction and final prediction [26, 27]. However, emotions flow naturally in the contextual sentences, and the sequential nature is nonnegligible. Flatten context processing makes the sequence of words too long and ignores the time step. It destroyed the hierarchical structural information of contextual sentences.

The hierarchical structure is a natural characteristic of text: words form sentences and sentences from contexts. This structure knowledge can be incorporated into model architecture for better representation. Inspired by this fact, some works try to stack deep learning architectures to provide specialized understanding at each level [28]. By hierarchical context modeling, each sentence is embedded into a vector representation by sentence-level encoder, and context information is further extracted by hierarchy context encoder [29]. Hierarchical Attention Networks (HANs) is proposed in [8], which mirrors the documents’ hierarchical structure. HANs mainly consists of two layers applied at word and sentence level, respectively. In each layer, there is a GRU-based sequence encoder along with an attention layer, which is aimed at paying more attention to important words and sentences that benefit to the final prediction.

Assume that we have N training sample X along with their contextual information C. Each sample x\in X is often a sentence from the hierarchical text, such as dialogue or blog, and its contextual information c=\left\{c_1,c_2,\cdots c_n\right\}\in C often means the preceding n sentences appeared before x. Each sample x is annotated with K emotional labels: \left\{e_1,\cdots e_k,\cdots e_K\right\}, which is denoted as a one hot vector y=\left\{y_1,\cdots y_k,\cdots y_K\right\}\in {ℝ}^{1\times K}, in which y_k=1is x contains emotion e_k otherwise y_k=0.

For each sample x\in X, a multilabel emotion recognition model F is trained to transform x into predicted distributions p=\left\{p_1,\cdots p_k,\cdots p_K\right\} based on its contextual information c and then give a final prediction of all possible emotion labels. The function F is denoted as

To model a sentence x along with its contextual information c=\left\{c_1,c_2,\cdots c_n\right\}, the simplest way is to utilizing flatten context modeling, by which x and contextual sentence c are concatenated as x^{\prime }=\left\{c_1,c_2,\cdots c_n,x\right\}, and all tokens in x^{\prime } are flattened into a word sequence. However, emotions flow naturally in each sentence. Such flatten processing makes the sequence of words too long and ignores the time step, which destroyed the hierarchical structural information. The sequential nature of context is nonnegligible, and such hierarchical information could better contribute to the emotion prediction.

Motivated by Hierarchical Attention Networks (HANs), we focus on hierarchical context modeling. Each sentence in x^{\prime }=\left\{c_1,c_2,\cdots c_n,x,\right\} is first encoded into sentence-level representation h^s=\left\{h_{c1}^s,h_{c2}^s,\cdots h_{cn}^s,h_x^s\right\} by a sentence-level encoder {En}_s, and then, contextual information is further encoded by a hierarchy context encoder. The framework of proposed hierarchical network with label embedding is shown in Figure 1.

For multilabel emotion recognition task, the training objective is often based on binary cross-entropy (BCE). However, BCE loss function takes each emotion as an independent individual and does not consider their relationships. Emotion correlation plays an essential role in this task, which makes emotion recognition be a more complex problem than traditional text classification. To guide the model to learn the emotion correlation during the training process, we propose an assembled training objective to consider all aspects.

The experiments are conducted on Chinese emotion corpus RenCECps (http://a1-www.is.tokushima-u.ac.jp/member/ren/Ren-CECps1.0/DocumentforRen-CECps1.0.html) to evaluate the proposed architecture. RenCECps is an annotated emotional corpus with Chinese blog texts. The corpus is annotated in document, paragraph, and sentence level [31]. Each level is annotated with eight emotional categories (“Joy,” “Hate,” “Love,” “Sorrow,” “Anxiety,” “Surprise,” “Anger,” and “Expect”).

Our experiments are conducted at sentence level, and the preceding two sentences of the current sentence are taken as the context information. After preprocessing, there are 24310 contextual sentences in training data and 6746 in testing data.

For each emotion e_i, its cumulative number {CN}_i is calculated.in which y_{n,i} is the annotation of emotion e_i in the n\text{th} sample, and the statistical results are shown in Figure 2. Label cardinality (LCard) is a standard measure of multilabeledness and means the average number of emotions concluded per sentence of the corpus and [32]. In RenCECps, LCard is 1.4468.

The global performance is evaluated by micro- and macro-F1 score. F1 score is the harmonic mean of precision and recall. Micro-F1 score gives each sample the same importance, while macro-F1 score takes all classes as equally important.

Some popular evaluation measures typically utilized in multilabel classification are utilized to measure the efficiency of proposed methods. Hamming Loss (HL) is the fraction of labels that are incorrectly predicted. Coverage evaluates how far it is needed to go down the ranked emotion list to cover all the relevant emotions in the instance. One Error (OE) evaluates the fraction of sentences whose top-ranked emotion is not in the relevant emotion set. Ranking Loss (RL) evaluates the average fraction of label pairs that are reversely ordered for instance.

For a given sentence, its preceding two sentences are taken as contextual sentences. There are total 8 emotion labels annotated for each sentence, and the dimension of label embedding is set to 256. The dimension of hidden state of GRU cell is set to 768/2, and 768 is the dimension of sentence-level embedding.

During the model training, the learning rate is set to 2e-5, and the batch size is set to 128. Adam optimization method is applied to train the model by minimizing the proposed training objective.

In this section, we report the experimental results of our proposed method and baseline models. Additionally, we analyze the influence of training objectives on output layer and label embedding.

We compare our proposed model with six baseline methods as follows.

(1) RERc [18]: a novel framework based on relevant emotion ranking to identify multiple emotions and produce the rankings of relevant emotions from text.

(2) HANs [8]: it has a hierarchical structure that mirrors the hierarchical structure of documents and has two levels of attention mechanisms applied at the word-and sentence-level. In our experiments, sentence-level encoder of HANs is replaced by pretrained BERT model.

(3) EDL [33]: Emotion Distribution Learning, it learns a mapping function from texts to their emotion distributions describing multiple emotions and their respective intensities based on label distribution learning.

(4) EmoDetect [34]: it outputs the emotion distribution based on a dimensionality reduction method using nonnegative matrix factorization, which combines several constraints such as emotions bindings, topic correlations, and emotion lexicons in a constraint optimization framework.

(5) ML-KNN [35]: Multi-Label k-Nearest Neighbor, which adapts the traditional k-nearest neighbor (KNN) algorithm to deal with multilabel data.

(6) Rank-SVM [36]: adapts maximum margin strategy to deal with multilabel data and focuses on distinguishing relevant from irrelevant while neglecting the rankings of relevant ones.

The comparison experiments of baseline HANs are implemented based on the open-source codes shared on GitHub, and the results of other baselines are adopted from the published papers.

The experimental results of our model compared with the baselines on the RenCECps dataset are shown in Table 1. The best result on each metric is in italics.

As the results shown in Table 1, it indicates that our proposed method significantly outperforms baseline models to a great extent and achieves satisfactory performance on RenCECps dataset. For example, compared to the baseline RERc, our model achieves an improvement of 10.73% micro-F1 score. On multilabel evaluation measures, our model achieves a reduction of 46.15% ranking loss and 21.78% one error. Compared to other baselines, our model achieved satisfactory results as well, which demonstrated the effectiveness of the proposed method.

Our proposed model is an extension of the baseline of HANs. In our experiments, sentence-level encoder of HANs is replaced by pretrained BERT model. Therefore, by comparing the results of these two models, it can be revealed whether the addition of label embedding layer is effective on the subtask of emotion correlation learning.

As we can see from the results shown in Table 1, the proposed model significantly outperforms baseline HANs, which achieves the improvement of micro-F1 score from 0.5573 to 0.5665 and macro-F1 score from 0.4003 to 0.4186. On multilabel evaluation measures, our model achieves a reduction of ranking loss from 0.1136 to 0.1131, one error from 0.3623 to 0.3559, and hamming loss from 0.2075 to 0.1998.

Both the proposed method and baseline HANs give prediction based on the contextual representation learned from a hierarchical network. HANs directly fed it into output layer for final prediction, which mainly consists of a fully connected layer and an activate function like sigmoid. This implementation is intuitive and straightforward, and it is also a common processing method in most multilabel classification tasks. However, such implementation treats emotion recognition task as a general text classification task. It does not consider the correlation between emotion labels, such as the probability of cooccurrence of “Love” and “Happy” is higher than that of “Love” and “Sad.” In our proposed model, label embedding space is introduced for emotion correlations learning. The final prediction is based on the interaction of the emotion representation of input text and the label embedding matrix. To guarantee that the label embedding matrix learned the semantic features among labels, training objective on label embedding is utilized to guide the training. The predicted probability given by label embedding matrix, as Equation (13), is visualized in Figure 3. The results in the figure clearly show that the label embedding matrix can accurately predict the corresponding emotion, which means that the emotional information of each label has been actually learned in the label embedding matrix.

As described in Section 3.3, we proposed an assembled training objective to realize the joint learning of both emotion recognition task and label embedding task. To evaluate the effectiveness of training objectives and label embedding network, we train the proposed model with different training objectives. The results are shown in Table 2. The symbols “M,” “C,” and “L” denote the loss function of multilabel loss, as in Equation (10), cosine embedding loss, as in Equation (11), and label embedding loss, as in Equation (14), which are utilized for training.

From Table 2, compared with the assembled training objective (“M+C+L”), the proposed model with only multilabel loss (“M”) on output layer achieves a reduction of 2.22% micro-F1 and 1.39% macro-F1 and an improvement of 12.37% ranking loss, 6.41% one-error, and 4.52% coverage. It suggests that the proposed ensemble training objective can contribute to the classification improvement.

The experimental results of the proposed model trained on “M+C” and “M” indicate the contribution of cosine embedding loss on the training of the label embedding matrix. Cosine embedding loss guides the training of label embedding by making the emotion representation of input being closer to the embedding of positive emotion labels while farther to other negative emotion labels.

The comparison results of the proposed model trained on “M+L” and “M” indicate that the addition of label embedding loss is effective. Label embedding loss guarantees that the trained label embedding matrix is able to encode semantic features among emotion labels.