Multi-task dialog act and sentiment recognition on Mastodon

Proceedings of the 27th International Conference on Computational Linguistics, pages 745–754 Santa Fe, New Mexico, USA, August 20-26, 2018. 745 Multi-task dialog act and sentiment recognition on Mastodon Christophe Cerisara, Somayeh Jafaritazehjani, Adedayo Oluokun and Hoa T. Le Universit´e de Lorraine, CNRS, LORIA, F-54000 Nancy, France cerisara@loria.fr, hoa.le@loria.fr Abstract Because of license restrictions, it often becomes impossible to strictly reproduce most research results on Twitter data already a few months after the creation of the corpus. This situation worsened gradually as time passes and tweets become inaccessible. This is a critical issue for reproducible and accountable research on social media. We partly solve this challenge by an- notating a new Twitter-like corpus from an alternative large social medium with licenses that are compatible with reproducible experiments: Mastodon. We manually annotate both dialogues and sentiments on this corpus, and train a multi-task hierarchical recurrent network on joint sen- timent and dialog act recognition. We experimentally demonstrate that transfer learning may be efficiently achieved between both tasks, and further analyze some specific correlations be- tween sentiments and dialogues on social media. Both the annotated corpus and deep network are released with an open-source license. 1 Introduction Social media are a gold mine for researchers in many domains and especially in natural language pro- cessing, because of the endless stream of linguistic content produced every day. However, a major issue faced by every researcher working with Twitter data concerns the accessibility of datasets previously extracted. Indeed, license restrictions limit the possibility to store tweets in a database for a long period of time, and the proportion of tweets that are continuously deleted from the Twitter company servers makes any Twitter corpus quickly obsolete and impossible to retrieve after a few months. This is a very serious issue for the ethics of experimental research, of which reproducibility is a foundational feature. Despite these limitations, many research and development works use Twitter data for a wide variety of applications. We propose in this work to exploit another social medium with language data, the Mastodon network, which closely resembles Twitter, except for two important differences: Mastodon is a decentralized social network, and it adopts permissive licenses that are compatible with reproducible research. In particular, everyone may create his own Mastodon server with his cohort of registered users, and have his server automatically federated within the worldwide social network. To enable this, the Mastodon software is released with the open-source AGPL license, and the user-generated content in the Mastodon network typically follows a Creative Commons license1, which allows redistribution of posts. Mastodon is very similar to Twitter with regard to content and usage, except that the user posts are limited by default to 500 characters, which gives the user the possibility to write longer sentences than on Twitter. From a sociological point of view, Mastodon users are mostly composed of people who are either attracted by the technical challenge of setting up their own server, or who have been disappointed by Twitter, sometimes because of advertisements, changes …

Content truncated to 500 words