Developing a twitter bot that can join a discussion using state-of-the-art architectures

Thesis Type: Postgraduate

Institution Of The Thesis: Middle East Technical University, Faculty of Engineering, Department of Computer Engineering, Turkey

Approval Date: 2019

Student: Yusuf Mücahit Çetinkaya



Twitter is today mostly used for sharing and commenting about news. In this manner, the interaction between Twitter users is inevitable. This interaction sometimes causes people to move daily debates to this social platform. Since being dominant in these debates is crucial, automation of this process becomes highly popular. In this work, we aim to train a bot that classifies posted tweets according to their semantic and generates logical tweets about a popular discussion, namely gun debate of the U.S. for this study. Bots are trained to tweet independently on their side of the debate and also reply to a tweet from opposite view. State-of-art architectures are tested to get more accurate classification. We have applied GloVe embedding model for representing tweets. Instead of using handcrafted features, long-short-term memory neural network is applied to these embeddings to get more informative and equal size feature vectors. This model is trained to encode the tweet by fed as a sequence of embeddings. Encoding is used for both classification and generation tasks. LSTM sequence to sequence model is used to generate tweets and replies to tweets. The attention mechanism is added to the reply model to produce more related tweets. We propose a new metric for measuring the relatedness of the reply to the target tweet. Additionally, human evaluators measure the quality of generated tweets according to relatedness to the topic and target tweet, which is replied.