Multi-Task Dialog Act and Sentiment Recognition on Mastodon

Last updated Jul 11, 2023

Release of 800k annotated posts (dialogue and sentiment) from single Mastodon instance.
Show limited transfer learning possible between both tasks though not correlated globally

# Abstract

Issue: Reproducibility when working on Twitter data
- license restrictions: cannot store tweets in a database for a long period
- proportion of tweets are continuously deleted
Solution: Use Mastodon data
- user-generated content typically follows CC licence
Comparison: Mastodon
- decentralized SN
- more characters per post
- mostly composed of people who are technical, disappointed by Twitter, changes of policies, privacy threads, belonging to minority group
- more international (most servers located in Japan and in the West of Europe)
- less developed than Twitter, but growing and useful for research
Focus of paper: Dialog acts and sentiment recognition
- Provide new Mastodon corpus
- With annotated sentiment
- Analyze dialog acts (speech acts)
- Show correlation between sentiment and dialog acts (via RNN transfer learning)