Social media plays a major role in everyday communication. While images andvideos are common in social media sites such as Facebook and Twitter , the text is still dominating the communication. Communication through text may lack non-verbal cues, and emojis can provide richer expression to mitigate this issue. Emojis are a set of reserved characters that are rendered as small pictograms that depict a facial expression [1,11]. In social media, sarcasm represents the nuanced form of language that individuals state the opposite of what is implied. Sarcasm detection is an important task to improve the quality of online communication. First, it helps us to understand the real intention of the user’s feedback. For example, user reviews can contain examples such as ‘Wow this product is great”, “It is very fast”, “Totally worth it”, etc. These comments, however, are being said in a sarcastic tone. Second, sarcastic posts may influence people’s emotions and reactions to the political campaign . The majority of existing sarcasm detection algorithms focuses on text information .
These include identifying the traits of the user from their past activities, responses texts, etc. Most of them have tried to train deep neural network models using the text to analyze sarcasm.
To overcome the challenges faced by all of these methods and for better performance, the Emoji can be considered to detect sarcasm. Emojis help us to find the tone of speech, the mood of the user and identify sarcasm in a better way.
Our contributions are summarized as follows:
– We provide a principled way to model emoji signals for social media post;
– We propose a new framework ESD which integrates text and emoji signals
into a coherent model for sarcasm detection; and
– We conduct experiments on real-world datasets to demonstrate the effectiveness
of the proposed framework ESD.
Preliminary Analysis Of Emoji Usage
Emojis serve as a medium for us to express certain opinions that can’t be expressed by our voice or body language. Emojis are the major contributing factor to the improvement in accuracy of our model because the neural network learns the connection between text and emojis. This analysis is performed to research in depth about the types of emojis used across the comments in the Twitter and Facebook data set. This gives us a clear picture of the most frequently used emojis in both sarcastic as well as non-sarcastic comments which in turn helps us to rank emojis based on their count of occurrences in the comments. The top 20 emojis used in our Twitter/Facebook data are visualized through the graphs. The following insights are obtained from the graphs.
– On comparison of emojis used across entire Facebook and Twitter data, the usage of Face with tongue out emoji is the highest (2.7K) among the sarcastic comments. The Face with tears of Joy, Loud crying face (2.6K), Grinning and Pouting face are the three specific emojis that are most frequently used with non-sarcastic comments.
Fig. 1. Comparison of top 20 Emojis for Facebook data.
Fig. 2. Comparison of top 20 Emojis for Twitter data.
– The number of other emojis used in sarcastic comments like winking face, the smirking face is found to be uniformly distributed across the Twitter data whereas emojis such as the loud crying face, pouting face, the confused face is observed to be uniformly distributed for the Facebook data.
– The usage of Face with stuck out tongue emoji is the first highest for Facebook data and third highest for Twitter data. However, the face with tears of joy emoji is being increasingly used in both sarcastic and non-sarcastic comments across the platforms.
– It is also clearly observed that the amount of Face with tongue out emoji in sarcastic comments is very high which is nearly times its usage in nonsarcastic comments for Facebook data. For twitter non-sarcastic comments, the count of this emoji is in-fact zero. This proves the fact that most of the comments having this emoji are clearly being sarcastic in nature.
In this section, we introduce the details of the proposed framework ESD for sarcasm detection on social media. It mainly consists of three components (see Figure 3) a text encoder, an emoji encoder, and a sarcasm prediction component. In general, the text encoder describes the mapping of words to latent representations; the emoji encoder illustrates the extraction of emoji latent representations, and the sarcasm prediction component learns a classification function to predict sarcasm in social media posts.
Fig. 3. The proposed framework ESD for sarcasm detection takes a list of words[?1,?2...??]and emojis[?1,?2,..??]as input and converts them into word[ ̃?1, ̃?2,... ̃??]and emoji[ ̃?1, ̃?2,... ̃??]embeddings.[ℎ1,ℎ2,...ℎ?]denotes the list of concatenated vectorswhich are passed through the bi-directional GRU. The attention weights[?1,?2,....??]are then multiplied and summed with the vector representations to give the contextvector,?. This vector?is finally passed to the sigmoid function for classification.
Comparison of Sarcasm Detection Methods
The representative state-of-the-art sarcasm detection methods that are compared with ESD, are listed as follows:
– FSNN : FSNN stands for Fracking Sarcasm using a Neural Networks, which uses a Convolutional Neural Network (CNN) followed by an LSTM and a Deep Neural Network (DNN) to detect sarcasm in a sentence.
– CASCADE : CASCADE stands for Contextual Sarcasm Detection in Online Discussion Forums. CASCADE uses CNNs to capture the user’s personality features to boost the performance of classification.
– RCCSD : RCCSD stands for The Role of Conversation Context for Sarcasm Detection, which uses conditional LSTM networks with sentence-level attention on conversational context and response.
Table 1. Best performance comparison for Sarcasm detection
1. Barbieri, F., Kruszewski, G., Ronzano, F., Saggion, H.: How cosmopolitan are emojis?: Exploring emojis usage and meaning over different languages with distributional semantics. In: Proceedings of the 2016 ACM on Multimedia Conference. pp. –. ACM (2016)
2. Eisner, B., Rocktäschel, T., Augenstein, I., Bošnjak, M., Riedel, S.: emojivec: Learning emoji representations from their description. arXiv preprint arXiv:1609.08359(2016)
3. Ghosh, A., Veale, T.: Fracking sarcasm using neural network. In: Proceedings of the 7th workshop on computational approaches to subjectivity, sentiment and social media analysis. pp. 161–169 (2016)
4. Ghosh, D., Fabbri, A.R., Muresan, S.: The role of conversation context for sarcasm detection in online interactions. arXiv preprint arXiv:1707.06226 (2017)
5. Hallsmar, F., Palm, J.: Multi-class Sentiment Classification on Twitter using an Emoji Training Heuristic. Master’s thesis, KTH Royal Institute of Technology School of Computer Science and Communication (05 2016)
6. Hazarika, D., Poria, S., Gorantla, S., Cambria, E., Zimmermann, R., Mihalcea, R.:Cascade: Contextual sarcasm detection in online discussion forums. arXiv preprint arXiv:1805.06413 (2018)
7. Hu, X., Tang, J., Gao, H., Liu, H.: Unsupervised sentiment analysis with emotional signals. In: Proceedings of the nd international conference on World Wide Web. pp. 607–618. ACM (2013)
8. Joshi, A., Bhattacharyya, P., Carman, M.J.: Automatic sarcasm detection: A survey. ACM Computing Surveys (CSUR) 50(5), 73(2017)
9. Kelly, R., Watts, L.: Characterising the inventive appropriation of emoji as relationally meaningful in mediated close personal relationships (2015)
10. Lee, H., Kwak, N.: The affect effect of political satire: Sarcastic humor, negative emotions, and political participation. Mass Communication and Society 17(3), 307–328 (2014)
11. Morstatter, F., Shu, K., Wang, S., Liu, H.: Cross-platform emoji interpretation: analysis, a solution, and applications. arXiv preprint arXiv:1709.04969 (2017)
12. Novak, P.K., Smailovic, J., Sluban, B., Mozetic, I.: Sentiment of emojis. PloS one 10(12), e0144296 (2015)
13. Pennington, J., Socher, R., Manning, C.D.: Glove: Global vectors for word representation. In: EMNLP. vol. 14, pp. 1532–43 (2014)
14. Rajadesingan, A., Zafarani, R., Liu, H.: Sarcasm detection on twitter: A behavioral modeling approach. In: Proceedings of the Eighth ACM International Conference on Web Search and Data Mining. pp. 97–106. ACM (2015)