Experiments on Fine Tuning Deep Learning Models With News Data For Tweet Classification
İbrahim Rıza Hallaç, Betül Ay, Galip Aydın · 2018
In this paper we present our experiments on tweet classification task. Because Twitter data is very noisy by its nature for including URLs, symbols, typos etc. it needs special treatment of text classification approaches. Also, availability of annotated Twitter data is rare except for sentiment analysis use cases. In this work we perform text classification on tweets for identifying whether a tweet belongs to one of the topics of culture, economy, politics, sports, or technology. Our approach shows that in the circumstance of having a very small amount of labeled tweet data we can classify tweets on high accuracy levels. To do this we first train a simple neural network with huge amount of news data. Then we apply basic fine-tuning steps on these models by using tweet data.