Argument Mining using multitask learning

Summary

The task of Argument mining is to extract automatically the arguments from texts, and essentially through that, the acquisition of a deeper understanding of language and logical meaning in it. It has several applications like decision making, text summarization, fact checking, even financial market prediction and inference problems in general. In this thesis we investigate how Multitask learning can help the task of Argument mining. Multitask learning is a type of transfer learning where we use different but related (auxiliary) tasks which are trained simultaneously with our main task. The motivation for using multitask learning is to bring knowledge from other tasks enriching the available knowledge during the training of the main task we want to solve. Our main task was the argument segmentation, on a dataset with persuasive essays. In our approach we explored the effect of various parameters in our models, such as the hierarchy of the architecture e.g. flat vs hierarchical, the choice of embeddings e.g. Glove vs contextualized embeddings, the effect of building our models on top of Bert (a very popular language model), and last but not least the effect of the type of task e.g. syntactic vs semantic as our auxiliary tasks. Specifically we found that tasks like POS-tagging and chunking did not showed good correlation with our main task of argument segmentation, and semantic tasks like NLI, or argument mining sub-tasks, showed better results. We also used a state of the art on Glue benchmark, multitask learning framework, the Mt-dnn, and we created a new set of argument mining tasks, based on the corresponding argument mining datasets we collected, which we used together with a number of argument mining sub-tasks (created in different degrees of granularity), based on our dataset of persuasive essays. For all pairs of the tasks, we examined which of them showed successful results.