We propose a system based on a densely connected LSTM network with multi-task learning strategy for the SemEval-2018 task. In our dense LSTM model, each layer takes all outputs from previous layers as input.
Rule-based methods depend on lexicons to identify irony, and traditional ML methods like SVM need manual feature enginnering.
SemEval 2018 aims to detect irony of tweets without explicit irony.
The embedding layer is used to convert the input tweets into a sequence of dense vectors. The POS tag features are one-hot encoded and concatenated with the word embedding vectors.
Since the irony information is complex, jointly using all levels of information is beneficial to predict irony more accurately.
The last layer outputs the hidden representation of texts, which can be concatenated with the sentiment features and the sentence embedding features. 3 dense layers with ReLU activation are used to predict for 3 different tasks: determining the missing ironic hashtag, identifying ironic or non-ironic, and identifying the irony types (verbal irony, other types of verbal irony, and situational irony).
By using this multi-task learning method, our model can incorporate different information such as the irony hashtags. In addition, classifying ironic/non-ironic and the irony types are similar tasks. Therefore, the performance of both tasks can be improved by combining the information of both tasks.
Since the layers in our Dense-LSTM can learn from all previous outputs, our model can combine different levels of contextual information to capture the high-level irony clues.