Increasingly, social media are used to express opinions and sentiments about companies, institutions, products, services, politics, etc. Companies and organizations in general have shown their interest in knowing the opinions and sentiments that their activities and products arouse in society. In this sense, Twitter has become an excellent tool for these purposes. The joint use of machine learning and natural language processing techniques allows us to automatically determine the opinion on a certain topic by analyzing social media. However, the performances of these sentiment analysis systems are clearly influenced by the use of figurative language such as irony. Determining whether a tweet is ironic or not is important as it generally changes the polarity of the tweet since the content of the message should not be interpreted literally. Our system treats this phenomenon for the English language using embedding adapted to this domain and language.

The main characteristic of our approach is that it uses deep learning models specifically trained for the English language. Concretely, it is based on the embeddings contextualization using the Transformers architecture. In addition, our system for irony detection for the English language has been tested in the SemEval2018 international competitions where it obtained the second-best results in the competition. To use EN-IRONIC, a Docker container is provided that allows it to run on Linux, Windows, and MacOS operating systems. The classification models are accessible through a web application independent of the operating system thanks to the provided REST API. The set of libraries and third-party software are freely distributed.