Review of Tweet Sarcasm Detection Using Deep Neural Network

Source: Deep Learning on Medium


With the current trend of deep neural networks outperforming most other learning models, it is no surprise that these are being used for detection of sarcasm with considerable success. In “Tweet Sarcasm Detection Using Deep Neural Network” Zhang et al., compares and analyses a deep neural network against more modest methods involving manual feature extraction.

Their proposed neural network uses a bi-directional gated recurrent neural network (GRNN) for capturing syntactic and semantic information about the tweets, and a pooling neural network to extract contextual features from historical tweets of the author.

To compare the effectiveness of this models, the authors used a discrete model classification model. This model consisted of a “local component” which extracted features such as word bigrams, trigrams from the tweet itself and a “contextual component” which consisted of features from author’s historical tweets. The two models are illustrated below:

The major distinction here is that, the feature extraction was ‘learned’ and done automatically in the neural network model, whereas features in the discrete model were extracted using manual implementations.

The authors used a tweet data-set consisting of manually annotated 9,104 sarcastic tweets. They then used twitter API for additional extraction of tweets concerning the tweet authors. The authors first analyzed the effect of GloVe word embedding and the results showed a sharp increase in performance for both models. As a final result, the authors obtained an accuracy of 79.29% and 78.55% using local tweets for neural network model and discrete model, respectively. Addition of contextual features yielded 90.74% accuracy for the neural network model — significantly higher than the discrete model.

For further reading please refer to Zhang, M., Zhang, Y., Guohong, F., 2016, “Tweet Sarcasm Detection Using Deep Neural Network.” pp. 2449–2460 In: Proc. of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers.