HiGRU: Hierarchical Gated Recurrent Units forUtterance-level Emotion Recognition

HiGRU: Hierarchical Gated Recurrent Units for
Utterance-level Emotion Recognition

Wenxiang Jiao, Haiqin Yang, Irwin King, and Michael R. Lyu
 Department of Computer Science and Engineering,
The Chinese University of Hong Kong, HKSAR, China
 Meitu and  The Hang Seng University of Hong Kong, HKSAR, China
{wxjiao,king,lyu}@cse.cuhk.edu.hk, hqyang@ieee.org
Abstract

In this paper, we address three challenges in utterance-level emotion recognition in dialogue systems: (1) the same word can deliver different emotions in different contexts; (2) some emotions are rarely seen in general dialogues; (3) long-range contextual information is hard to be effectively captured. We therefore propose a hierarchical Gated Recurrent Unit (HiGRU) framework with a lower-level GRU to model the word-level inputs and an upper-level GRU to capture the contexts of utterance-level embeddings. Moreover, we promote the framework to two variants, HiGRU with individual features fusion (HiGRU-f) and HiGRU with self-attention and features fusion (HiGRU-sf), so that the word/utterance-level individual inputs and the long-range contextual information can be sufficiently utilized. Experiments on three dialogue emotion datasets, IEMOCAP, Friends, and EmotionPush demonstrate that our proposed HiGRU models attain at least 8.7%, 7.5%, 6.0% improvement over the state-of-the-art methods on each dataset, respectively. Particularly, by utilizing only the textual feature in IEMOCAP, our HiGRU models gain at least 3.8% improvement over the state-of-the-art conversational memory network (CMN) with the trimodal features of text, video, and audio.

\aclfinalcopy

1 Introduction

Emotion recognition is a significant artificial intelligence research topic due to the promising potential of developing empathetic machines for people. Emotion is a universal phenomena across different cultures and mainly consists of six basic types: anger, disgust, fear, happiness, sadness, and surprise Ekman (1971, 1992).

Figure 1: The word “okay” exhibits different emotions in the American television sitcom, Friends.

In this paper, we focus on textual dialogue systems because textual feature dominates the performance over audio and video features Poria et al. (2015, 2017). In utterance-level emotion recognition, an utterance Olson (1977) is a unit of speech bounded by breathes or pauses and its goal is to tag each utterance in a dialogue with the indicated emotion.

In this task, we address three challenges: First, the same word can deliver different emotions in different contexts. For example, in Figure 1, the word “okay” can deliver three different emotions, anger, neutral, and joy, respectively. Strong emotions like joy and anger may be indicated by the symbols “!” or “?” along the word. To identify a speaker’s emotion precisely, we need to explore the dialogue context sufficiently. Second, some emotions are rarely seen in general dialogues. For example, people are usually calm and present a neutral emotion while only in some particular situations, they express strong emotions, like anger or fear. Thus we need to be sensitive to the minority emotions while relieving the effect of the majority emotions. Third, the long-range contextual information is hard to be effectively captured in an utterance/dialogue, especially when the length of an utterance/dialogue in the testing set is longer than those in the training set.

To tackle these challenges, we propose a hierarchical Gated Recurrent Unit (HiGRU) framework for the utterance-level emotion recognition in dialogue systems. More specifically, HiGRU is composed by two levels of bidirectional GRUs, a lower-level GRU to model the word sequences of each utterance to produce individual utterance embeddings, and an upper-level GRU to capture the sequential and contextual relationship of utterances. We further promote the proposed HiGRU to two variants, HiGRU with individual features fusion (HiGRU-f), and HiGRU with self-attention and features fusion (HiGRU-sf). In HiGRU-f, the individual inputs, i.e., the word embeddings in the lower-level GRU and the individual utterance embeddings in the upper-level GRU, are concatenated with the hidden states to generate the contextual word/utterance embeddings, respectively. In HiGRU-sf, a self-attention layer is placed on the hidden states from the GRU to learn long-range contextual embeddings, which are concatenated with the original individual embeddings and the hidden states to generate the contextual word/utterance embeddings. Finally, the contextual utterance embedding is sent to a fully-connected (FC) layer to determine the corresponding emotion. To alleviate the effect of data imbalance issue, we follow Khosla (2018) to train our models by minimizing a weighted categorical cross-entropy.

We summarize our contributions as follows:

  • We propose a HiGRU framework to better learn both the individual utterance embeddings and the contextual information of utterances, so as to recognize the emotions more precisely.

  • We propose two progressive HiGRU variants, HiGRU-f and HiGRU-sf, to sufficiently incorporate the individual word/utterance-level information and the long-range contextual information respectively.

  • We conduct extensive experiments on three textual dialogue emotion datasets, IEMOCAP, Friends, and EmotionPush. The results demonstrate that our proposed HiGRU models achieve at least 8.7%, 7.5%, 6.0% improvement over state-of-the-art methods on each dataset, respectively. Particularly, by utilizing only the textual feature in IEMOCAP, our proposed HiGRU models gain at least 3.8% improvement over the existing best model, conversational memory network (CMN) with not only the text feature, but also the visual, and audio features.

2 Related Work

Text-based emotion recognition is a long-standing research topic Wilson et al. (2004); Yang et al. (2007); Medhat et al. (2014). Nowadays, deep learning technologies have become dominant methods due to the outstanding performance. Some prominent models include recursive autoencoders (RAEs) Socher et al. (2011), convolutional neural networks (CNNs) Kim (2014), and recurrent neural networks (RNNs) Abdul-Mageed and Ungar (2017). However, these models treat texts independently thus cannot capture the inter-dependence of utterances in dialogues Kim (2014); Lai et al. (2015); Grave et al. (2017); Chen et al. (2016); Yang et al. (2016). To exploit the contextual information of utterances, researchers mainly explore in two directions: (1) extracting contextual information among utterances, or (2) enriching the information embedded in the representations of words and utterances.

Contextual Information Extraction.

The RNN architecture is a standard way to capture the sequential relationship of data. Poria et al. propose a bidirectional contextual long short-term memory (LSTM) network, termed bcLSTM, to model the context of textual features extracted by CNNs. Hazarika et al. improve bcLSTM by a conversational memory network (CMN) to capture the self and inter-speaker emotional influence, where GRU is utilized to model the self-influence and the attention mechanism is employed to excavate the inter-speaker emotional influence. Though CMN is reported to attain better performance than bcLSTM on IEMOCAP Hazarika et al. (2018), the memory network is too complicated for small-size dialogue datasets.

Representation Enrichment.

Multimodal features have been utilized to enrich the representation of utterances Poria et al. (2015, 2017). Previous work indicate that textual features dominate the performance of recognizing emotions in contrast to visual or audio features Poria et al. (2015, 2017). Recently, the textual features are mainly extracted by CNNs to learn individual utterance embeddings Poria et al. (2015, 2017); Zahiri and Choi (2018); Hazarika et al. (2018). However, CNNs do not capture the contextual information within each utterance well.

Figure 2: The architecture of our proposed HiGRU-sf. “Attention” denotes self-attention. By removing the “Attention” layer, we attain HiGRU-f, and by further removing the “Fusion” layer, we can recover the vanilla HiGRU.

On the other hand, hierarchical RNNs have been proposed and demonstrated good performance in conventional text classification task Tang et al. (2015), dialogue act classification Liu et al. (2017); Kumar et al. (2018), and speaker change detection Meng et al. (2017). But they are not well explored in the task of utterance-level emotion recognition in dialogue systems.

3 Approach

The task of utterance-level emotion recognition is defined as follows:

Definition 1 (Utterance-level Emotion Recognition).

Suppose we are given a set of dialogues, , where is the number of dialogues. In each dialogue, , is a sequence of utterances, where the utterance is spoken by the speaker with a certain emotion . All speakers compose the set and the set consists of all emotions, such as anger, joy, sadness, and neutral. Our goal is to train a model to tag each new utterance with an emotion label from as accurately as possible.

To solve this task, we propose a hierarchical Gated Recurrent Units (HiGRU) framework and extend two progressive variants, HiGRU with individual features fusion (HiGRU-f) and HiGRU with self-attention and features fusion (HiGRU-sf) (illustrated in Figure 2).

3.1 HiGRU: Hierarchical GRU

The vanilla HiGRU consists of two-level GRUs: the lower-level bidirectional GRU is to learn the individual utterance embedding by modeling the word sequence within an utterance and the upper-level bidirectional GRU is to learn the contextual utterance embedding by modeling the utterance sequence within a dialogue.

Individual Utterance Embedding.

For the utterance in , , where is the number of words in the utterance . The corresponding sequence of individual word embeddings are fed into the lower-level bidirectional GRU Cho et al. (2014) to learn the individual utterance embedding in two opposite directions:

(1)
(2)

The two hidden states and are concatenated into to produce the contextual word embedding for via the activation function on a linear transformation:

(3)

where and are the model parameters, and are the dimensions of word embeddings and the hidden states of the lower-level GRU, respectively.

The individual utterance embedding is then obtained by max-pooling on the contextual word embeddings within the utterance:

(4)

Contextual Utterance Embedding.

For the dialogue, , the learned individual utterance embeddings, , are fed into the upper-level bidirectional GRU to capture the sequential and contextual relationship of utterances in a dialogue:

(5)
(6)

Here, the hidden states of the upper-level GRU are represented by , to distinguish from those learned in the lower-level GRU denoted by . Accordingly, we can obtain the contextual utterance embedding by

(7)

where , and are the model parameters, is the dimension of the hidden states in the upper-level GRU. Since the emotions are recognized at utterance-level, the learned contextual utterance embedding is directly fed to a FC layer followed by a softmax function to determine the corresponding emotion label:

(8)

where is the predicted vector over all emotions, and , .

3.2 HiGRU-f: HiGRU + Individual Features Fusion

The vanilla HiGRU contains two main issues: (1) the individual word/utterance embeddings are diluted with the stacking of layers; (2) the upper-level GRU tends to gather more contextual information from the majority emotions, which deteriorates the overall model performance.

To resolve these two problems, we propose to fuse individual word/utterance embeddings with the hidden states from GRUs so as to strengthen the information of each word/utterance in its contextual embedding. This variant is named as HiGRU-f, representing HiGRU with individual features fusion. Hence, the lower-level GRU can maintain individual word embeddings and the upper-level GRU can relieve the effect of majority emotions and attain a more precise utterance representation for different emotions. Specifically, the contextual embeddings are updated as:

(9)
(10)

where , , , and .

3.3 HiGRU-sf: HiGRU + Self-Attention and Feature Fusion

Another challenging issue is to extract the contextual information of long sequences, especially the sequences in the testing set that are longer than those in the training set Bahdanau et al. (2014). To fully utilize the global contextual information, we place a self-attention layer upon the hidden states of HiGRU and fuse the attention outputs with the individual word/utterance embeddings and the hidden states to learn the contextual word/utterance embeddings. Hence, this variant is termed HiGRU-sf, representing HiGRU with self-attention and features fusion.

Particularly, we apply self-attention upon the forward and backward hidden states separately to produce the left context embedding, (), and the right context embedding, (), respectively. This allows us to gather the unique global contextual information at the current step in two opposite directions and yield the corresponding contextual embeddings computed as follows:

(11)
(12)

where , , , and .

Figure 3: Self-attention over the forward hidden states of GRU.

Self-Attention (SA).

The self-attention mechanism is an effective non-recurrent architecture to compute the relation between one input to all other inputs and has been successfully applied in various natural language processing applications such as reading comprehension Hu et al. (2018), and neural machine translation Vaswani et al. (2017). Figure 3 shows the dot-product SA over the forward hidden states of GRU to learn the left context . Each element in the attention matrix is computed by

(13)

An attention mask is then applied to waive the inner attention between the sequence inputs and paddings. At each step, the corresponding left context is then computed by the weighted sum of all the forward hidden states:

(14)

where is the weight of to be included in . The right context can be computed similarly.

Dataset #Dialogue (#Utterance) #Emotion
Train Val Test Ang Hap/Joy Sad Neu Others
IEMOCAP 96 (3,569) 24 (721) 31 (1,208) 1,090 1,627 1,077 1,704 0
Friends 720 (10,561) 80 (1,178) 200 (2,764) 759 1,710 498 6,530 5,006
EmotionPush 720 (10,733) 80 (1,202) 200 (2,807) 140 2,100 514 9,855 2,133
Table 1: Statistics of the textual dialogue datasets.

3.4 Model Training

Following Khosla (2018) which attains the best performance in the EmotionX shared task Hsu and Ku (2018), we minimize a weighted categorical cross-entropy on each utterance of all dialogues to optimize the model parameters:

(15)

where is the original one-hot vector of the emotion labels, and and are the elements of and corresponding to the class .

Similar to Khosla (2018), we assign the loss weight inversely proportional to the number of training utterances in the class , denoted by , i.e., assigning larger loss weights for the minority classes to relieve the data imbalance issue. The difference is that we add a constant to adjust the smoothness of the distribution. Then, we have:

(16)

4 Experiments

We conduct systematical experiments to demonstrate the advantages of our proposed HiGRU models.

4.1 Datasets

The experiments are carried out on three textual dialogue emotion datasets (see the statistics in Table 1):

Iemocap111https://sail.usc.edu/iemocap/.

It contains approximately 12 hours of audiovisual data, including video, speech, motion capture of face, text transcriptions. Following Poria et al. (2017); Hazarika et al. (2018): (1) We apply the first four sessions for training and the last session for test; (2) The validation set is extracted from the shuffled training set with the ratio of 80:20; (3) We only evaluate the performance on four emotions: anger, happiness, sadness, neutral, and remove the rest utterances.

Friends222http://doraemon.iis.sinica.edu.tw/emotionlines.

The dataset is annotated from the Friends TV Scripts Hsu and Ku (2018), where each dialogue in the dataset consists of a scene of multiple speakers. Totally, there are 1,000 dialogues, which are split into 720, 80, and 200 dialogues for training, validation, and testing, respectively. Each utterance in a dialogue is labeled by one of the eight emotions: anger, joy, sadness, neutral, surprise, disgust, fear, and non-neutral.

EmotionPush333http://doraemon.iis.sinica.edu.tw/emotionlines.

The dataset consists of private conversations between friends on the Facebook messenger collected by an App called EmotionPush, which is released for the EmotionX shared task Hsu and Ku (2018). Totally, there are 1,000 dialogues, which are split into 720, 80, 200 dialogue for training, validation, and testing, respectively. All the utterances are categorized into one of the eight emotions as in the Friends dataset.

Following the setup of Hsu and Ku (2018), in Friends and EmotionPush, we only evaluate the model performance on four emotions: anger, joy, sadness, and neutral, and we exclude the contribution of the rest emotion classes during training by setting their loss weights to zero.

Data Preprocessing.

We preprocess the datasets by the following steps: (1) The utterances are split into tokens with each word being made into the lowercase; (2) All non-alphanumerics except “?” and “!” are removed because these two symbols usually exhibit strong emotions, such as surprise, joy and anger; (3) We build a dictionary based on the words and symbols extracted, and follow  Poria et al. (2017) to represent the tokens by the publicly available 300-dimensional word2vec444https://code.google.com/archive/p/word2vec/ vectors trained on 100 billion words from Google News. The tokens not included in the word2vec dictionary are initialized by randomly-generated vectors.

4.2 Evaluation Metrics

To conduct fair comparison, we adopt two metrics as Hsu and Ku (2018), the weighted accuracy (WA) and unweighted accuracy (UWA):

(17)

where is the percentage of the class in the testing set, and is the corresponding accuracy.

Generally, recognizing strong emotions may provide more value than detecting the neutral emotion Hsu and Ku (2018). Thus, in Friends and EmotionPush, UWA is a more favorite evaluation metric because WA is heavily compromised with the large proportion of the neutral emotion.

4.3 Compared Methods

Our proposed vanilla HiGRU, HiGRU-f, and HiGRU-sf555https://github.com/wxjiao/HiGRUs are compared with the following state-of-the-art baselines:

bcLSTM Poria et al. (2017): a bidirectional contextual LSTM with multimodal features extracted by CNNs;

CMN Hazarika et al. (2018): a conversational memory network with multimodal features extracted by CNNs;

SA-BiLSTM Luo et al. (2018): a self-attentive bidirectional LSTM model, a neat model achieving the second place of EmotionX Challenge Hsu and Ku (2018);

CNN-DCNN Khosla (2018): a convolutional-deconvolutional autoencoder with more handmade features, the winner of EmotionX Challenge Hsu and Ku (2018);

bcLSTM and bcGRU: our implemented bcLSTM and bcGRU with the weighted loss on the textual feature extracted from CNNs.

4.4 Training Procedure

All our implementations are coded on the Pytorch framework. To prevent the models fitting the order of data, we randomly shuffle the training set at the beginning of every epoch.

Parameters.

For bcLSTM and bcGRU, the CNN layer follows the setup of Kim (2014), i.e., consisting of the kernels of 3, 4, and 5 with 100 feature maps each. The convolution results of each kernel are fed to a max-over-time pooling operation. The dimension of the hidden states of the upper-level bidirectional LSTM or GRU is set to 300. For HiGRU, HiGRU-f, and HiGRU-sf, the dimensions of hidden states are set to 300 for both levels. The final FC layer contains two sub-layers with 100 neurons each.

Training.

We adopt Adam Kingma and Ba (2014) as the optimizer and set an initial learning rate, for IEMOCAP and for Friends and EmotionPush, respectively. An annealing strategy is utilized by decaying the learning rate by half every 20 epochs. Early stopping with a patience of 10 is adopted to terminate training based on the accuracy of the validation set. Specifically, following the best models on each dataset, the parameters are tuned to optimize WA on the validation set of IEMOCAP and to optimize UWA on the validation set of Friends and EmotionPush, respectively. Gradient clipping with a norm of 5 is applied to model parameters. To prevent overfitting, dropout with a rate of 0.5 is applied after the contextual word/utterance embeddings, and the FC layer.

Loss weights.

For Friends and EmotionPush, as mentioned in Section 4.1, the loss weights are set to zero except the four considered emotions, to ignore the others during training. Besides, the power rate of loss weights is tested from 0 to 1.5 with a step of 0.25, and we use the best one for each model and dataset.

Model (Feat) Ang Hap Sad Neu WA UWA
bcLSTM1  (T) 76.07 78.97 76.23 67.44 73.6 74.6
(T+V+A) 77.98 79.31 78.30 69.92 76.1 76.3
CMN2  (T) - - - - 74.1 -
(T+V+A) 89.88 81.75 77.73 67.32 77.6 79.1
bcLSTM (T) 75.29 79.40 78.07 76.53 77.7(1.1) 77.3(1.4)
bcGRU (T) 77.20 80.99 76.26 72.50 76.9(1.6) 76.7(1.3)
HiGRU (T) 75.41 91.64 79.79 70.74 80.6(0.5) 79.4(0.5)
HiGRU-f (T) 76.69 88.91 80.25 75.92 81.5(0.7) 80.4(0.5)
HiGRU-sf (T) 74.78 89.65 80.50 77.58 82.1(0.4) 80.6(0.2)
  • by Poria et al. (2017);

  • by Hazarika et al. (2018).

Table 2: Experimental results on IEMOCAP. “(Feat)” represents the features used in the models, where T, V, and A denote the textual, visual, and audio features, respectively. The underlined results of bcLSTM and CMN are derived by us accordingly, while “-” means the results are unavailable from the original paper.
Model Train Friends (F) EmotionPush (E)
Ang Joy Sad Neu WA UWA Ang Joy Sad Neu WA UWA
SA-BiLSTM1 F+E 49.1 68.8 30.6 90.1 - 59.6 24.3 70.5 31.0 94.2 - 55.0
CNN-DCNN2 F+E 55.3 71.1 55.3 68.3 - 62.5 45.9 76.0 51.7 76.3 - 62.5
bcLSTM F(E) 64.7 69.6 48.0 75.6 72.4(4.2) 64.4(1.6) 32.9 69.9 47.1 78.0 74.7(4.4) 57.0(2.1)
bcGRU F(E) 69.5 65.4 52.9 74.7 71.7(4.7) 65.6(1.2) 33.7 71.1 57.2 76.1 73.9(2.9) 59.5(1.8)
bcLSTM F+E 54.5 75.6 43.4 73.0 70.5(4.5) 61.6(1.6) 52.4 79.1 54.7 73.3 73.4(3.8) 64.9(2.1)
bcGRU F+E 59.0 78.6 42.3 71.4 70.2(5.1) 62.8(1.4) 49.4 74.8 61.9 72.4 72.1(4.3) 64.6(1.8)
HiGRU F(E) 66.9 73.0 51.8 77.2 74.4(1.7) 67.2(0.6) 55.6 78.1 57.4 73.8 73.8(2.0) 66.3(1.7)
HiGRU-f F(E) 69.1 72.1 60.4 72.1 71.3(2.9) 68.4(1.0) 55.9 78.9 60.4 72.4 73.0(2.2) 66.9(1.2)
HiGRU-sf F(E) 70.7 70.9 57.7 76.2 74.0(1.4) 68.9(1.5) 57.5 78.4 64.1 72.5 73.0(1.6) 68.1(1.2)
HiGRU F+E 55.4 81.2 51.4 64.4 65.8(4.2) 63.1(1.5) 50.8 76.9 69.0 75.7 75.3(1.7) 68.1(1.2)
HiGRU-f F+E 54.9 78.3 55.5 68.7 68.5(3.0) 64.3(1.2) 58.3 79.1 69.6 70.0 71.5(2.5) 69.2(0.9)
HiGRU-sf F+E 56.8 81.4 52.2 68.7 69.0(2.0) 64.8(1.3) 57.8 79.3 66.3 77.4 77.1(1.0) 70.2(1.1)
  • by Luo et al. (2018);

  • by Khosla (2018).

Table 3: Experimental results on Friends and EmotionPush. In the Train column, F(E) denotes the model is trained on only one training set, Friends or EmotionPush. F+E means the model is trained on the mixed training set while validated and tested individually.

4.5 Main Results

Table 2 and Table 3 report the average results of 10 trials each on the three datasets, where the standard deviations of WA and UWA are recorded by the subscripts in round brackets. The results of bcLSTM, CMN, SA-BiLSTM, and CNN-DCNN are copied directly from the original papers for a fair comparison because we follow the same configuration for the corresponding datasets. From the results, we have the following observations:

(1) Baselines.

Our implemented bcLSTM and bcGRU, attain comparable performance with the state-of-the-art methods on all three datasets.

From the results on IEMOCAP in Table 2, we observe that: (a) By utilizing the textual feature only, bcGRU outperforms bcLSTM and CMN trained on the textual feature significantly, attaining +3.3 and +2.8 gain in terms of WA, respectively. bcLSTM performs better than bcGRU, and even beats bcLSTM and CMN with the trimodal features in terms of WA. In terms of UWA, CMN performs better than bcLSTM only when it is equipped with multimodal features. (b) By examining the detailed accuracy in each emotion, bcLSTM and bcGRU with the textual feature attain much higher accuracy on the neutral emotion than bcLSTM with the only textual feature while maintaining good performance on the other three emotions. The results show that the weighted loss function benefits the training of models.

From the results on Friends and EmotionPush in Table 3, we observe that bcLSTM and bcGRU trained on the same dataset (F+E) of CNN-DCNN perform better than CNN-DCNN on EmotionPush while attaining comparable performance with CNN-DCNN on Friends. The results show that by utilizing the contextual information with the weighted loss function, bcLSTM and bcGRU can beat the state-of-the-art method.

(2) HiGRUs vs. Baselines.

Our proposed HiGRUs outperform the state-of-the-art methods with significant margins on all the datasets.

From Table 2, we observe that: (a) CMN with the trimodal features attains the best performance on the anger emotion while our vanilla HiGRU achieves the best performance on the happiness emotion and gains further improvement on sadness and neutral emotions over CMN. Overall, the vanilla HiGRU achieves at least 8.7% and 3.8% improvement over CMN with the textual feature and the trimodal features in terms of WA, respectively. The results, including those of bcLSTM and bcGRU, indicate that GRU learns better representations of utterances than CNN in this task. (b) The two variants, HiGRU-f and HiGRU-sf, can further attain +0.9 and +1.5 improvement over HiGRU in terms of WA and +1.0 and +1.2 improvement over HiGRU in terms of UWA, respectively. The results demonstrate that the included individual word/utterance-level features and long-range contextual information in HiGRU-f and HiGRU-sf, are indeed capable of boosting the performance of the vanilla HiGRU.

From Table 3, we can see that: (a) In terms of UWA, HiGRU trained and tested on individual sets of Friends and EmotionPush gains at least 7.5% and 6.0% improvement over CNN-DCNN, respectively. Overall, our proposed HiGRU achieves well-balanced performance for the four tested emotions, especially attaining significant better performance on the minority emotions of anger and sadness. (b) Moreover, HiGRU-f and HiGRU-sf further improve HiGRU +1.2 accuracy and +1.7 accuracy on Friends and +0.6 accuracy and +1.8 accuracy on EmotionPush in terms of UWA, respectively. The results again demonstrate the superior power of HiGRU-f and HiGRU-sf.

(3) Mixing Training Sets.

By examining the results from the last ten rows in Table 3, we conclude that it does not necessarily improve the performance by mixing the two sets of training data.

Though the best performance of SA-BiLSTM and CNN-DCNN is obtained by training on the mixed dataset, the testing results show that our implemented bcLSTM, bcGRU and our proposed HiGRU models can attain better performance on EmotionPush but yield worse performance on Friends in terms of UWA.

By examining the detailed emotions, we speculate that: EmotionPush is a highly imbalanced dataset with over 60% of utterances in the neutral emotion. Introducing EmotionPush into a more balanced dataset, Friends, is equivalent to down-sampling the minority emotions in Friends. This hurts the performance on the minority emotions, anger and sadness. Meanwhile, introducing Friends into EmotionPush corresponds to up-sampling the minority emotions in EmotionPush. The performance of the sadness emotion is significantly boosted and that on the anger emotion is at least unaffected.

4.6 Discussions

bcGRU HiGRU HiGRU-f HiGRU-sf
- 65.6(1.2) - - -
300 - 67.2(0.6) 68.4(1.0) 68.9(1.5)
200 - 67.6(2.0) 68.9(0.9) 69.1(1.3)
150 - 67.6(1.5) 68.5(1.3) 68.9(1.2)
100 - 67.5(1.7) 68.4(1.3) 69.6(1.0)
Table 4: Experimental results of UWA on Friends by our proposed models with different scales of utterance encoder.

Model Size.

We study how the scale of the utterance encoder affects the performance of our proposed models, especially when our models contain a similar number of parameters as the baseline, say bcGRU. Such a fair condition can be made between our HiGRU-sf and bcGRU if we set to 150. From the testing results on Friends in Table 4, we can observe that: (1) Under the fair condition, the performance of our HiGRU-sf is not degraded compared to that when . HiGRU-sf still outperforms bcGRU by a significant margin. (2) Overall, no matter is larger or smaller than 150, HiGRU-sf maintains consistently good performance and the difference between HiGRU-sf and HiGRU-f or HiGRU keeps noticeable. These results further demonstrate the superiority of our proposed models over the baseline bcGRU and the motivation of developing the two variants based on the vanilla HiGRU.

Role    Utterance Truth bcGRU HiGRU-sf
Scene-1
Phoebe Okay. Oh but don’t tell them Monica’s pregnant because they frown on that. Neu Neu Neu
Rachel Okay. Neu Neu Neu
Phoebe Okay. Neu Neu Neu
Scene-2
Phoebe Yeah! Sure! Yep! Oh, y’know what? If I heard a shot right now, I’d throw my body on you. Joy Ang Joy
Gary Oh yeah? Well maybe you and I should take a walk through a bad neighborhood. Other / /
Phoebe Okay! Joy Ang Joy
Gary All right. Neu Neu Neu
Scene-3
Female Can I send you, like videos and stuff? What about when they start walking. Other / /
Male Yeah yeah yeah. Sad Hap Sad
Male You you record every second. You record every second because I want to see it all. Okay? Hap Hap Sad
Male If I don’t get to see it now, I get to see it later at least, you know? You’ve got to keep it all for me; all right? Other / /
Female Okay. Sad Neu Sad
Table 5: “Okay” expresses distinct emotions in three different scenes.

Successful Cases.

We investigate three scenes related to the word “okay” that expresses three distinct emotions. The first two scenes come from the testing set of Friends and the third one from that of IEMOCAP. We report the predictions made by bcGUR and our HiGRU-sf, respectively, in Table 5. In Scene-1, “okay” with period usually exhibits little emotion and both bcGRU and HiGRU-sf correctly classify it as “Neu”. In Scene-2, “okay” with “!” expresses strong emotion. However, bcGRU misclassifies it to “Ang” while HiGRU-sf successfully recognizes it as “Joy”. Actually, the mistake can be traced back to the first utterance of this scene which is also misclassified as “Ang”. This indicates that bcGRU tends to capture the wrong atmosphere within the dialogue. As for Scene-3, “okay” with period now indicates “Sad” and is correctly recognized by HiGRU-sf but misclassified as “Neu” by bcGRU. Note that HiGRU-sf also classifies the third utterance in Scene-3 as “Sad” which seems to be conflicting to the ground truth. In fact, our HiGRU-sf captures the blues of this parting situation, where the true label “Hap” may not be that suitable. These results show that our HiGRU-sf learns from both each utterance and the context, and can make correct predictions of the emotion of each utterance.

Role    Utterance Truth bcGRU HiGRU-sf
Scene-4
Ross Hi. Neu Neu Neu
Rachel Hi. Neu Neu Neu
Ross Guess what? Neu Neu Neu
Rachel What? Neu Neu Neu
Ross They published my paper. Joy Sad Neu
Rachel Oh, really, let me see, let me see. Joy Neu Neu
Phoebe Rach, look! Oh, hi! Where is my strong Ross Skywalker to come rescue me. There he is. Other / /
Scene-5
Speaker-1 Sorry for keeping you up Sad Sad Sad
Speaker-2 Lol don’t be Joy Joy Joy
Speaker-2 I didn’t have to get up today Neu Sad Sad
Speaker-1 :p Joy Joy Joy
Speaker-2 It’s actually been a really lax day Joy Neu Sad
Table 6: Wrong predictions made by both bcGRU and our HiGRU-sf in two scenes.

Failed Cases.

At last, we show some examples that both bcGRU and our HiGRU-sf fail in recognizing the right emotions in Table 6, i.e., Scene-4 from Friends and Scene-5 from EmotionPush. In Scene-4, both bcGRU and HiGRU-sf make wrong predictions for the fifth and the sixth utterances. It should be good news that Ross has his paper published and Rachel is glad to see related reports about it. However, the transcripts do not reveal very strong emotions compared to what the characters might act in the TV show. This kind of scenes may be addressed by incorporating some other features like audio and video. As for Scene-5, the third and the fifth utterances are classified into wrong emotions. Notice that the emotions indicated from the two utterances are very subtle even for humans. The Speaker-2 did not plan to get up today, but Speaker-1 kept him/her up and it ended up with a really lax day. So, the Speaker-2 feels joyful now. This indicates that even taking into the context into account, the models’ capability of understanding subtle emotions is still limited and more exploration is required.

5 Conclusion

We propose a hierarchical Gated Recurrent Unit (HiGRU) framework to tackle the utterance-level emotion recognition in dialogue systems, where the individual utterance embeddings are learned by the lower-level GRU and the contexts of utterances are captured by the upper-level GRU. We promote the HiGRU framework to two variants, HiGRU-f, and HiGRU-sf, and effectively capture the word/utterance-level inputs and the long-range contextual information, respectively. Experimental results demonstrate that our proposed HiGRU models can well handle the data imbalance issue and sufficiently capture the available text information, yielding significant performance boosting on all three tested datasets. In the future, we plan to explore semi-supervised learning methods to address the problem of data scarcity in this task.

Acknowledgments

This work is supported by the Research Grants Council of the Hong Kong Special Administrative Region, China (No. CUHK 14208815 and No. CUHK 14210717 of the General Research Fund, and Project No. UGC/IDS14/16), and Meitu (No. 7010445). We thank the three anonymous reviewers for the insightful suggestions on various aspects of this work.

References

Comments 0
Request Comment
You are adding the first comment!
How to quickly get a good reply:
  • Give credit where it’s due by listing out the positive aspects of a paper before getting into which changes should be made.
  • Be specific in your critique, and provide supporting evidence with appropriate references to substantiate general statements.
  • Your comment should inspire ideas to flow and help the author improves the paper.

The better we are at sharing our knowledge with each other, the faster we move forward.
""
The feedback must be of minimum 40 characters and the title a minimum of 5 characters
   
Add comment
Cancel
Loading ...
350875
This is a comment super asjknd jkasnjk adsnkj
Upvote
Downvote
""
The feedback must be of minumum 40 characters
The feedback must be of minumum 40 characters
Submit
Cancel

You are asking your first question!
How to quickly get a good answer:
  • Keep your question short and to the point
  • Check for grammar or spelling errors.
  • Phrase it like a question
Test
Test description