Generating argumentation frameworks from text

Oana Cocarascu, Francesca Toni
2020
Artificial Intelligence targets the design and creation of intelligent machines that behave like humans. One fundamental aspect of human intelligence is the ability to argue. In the past years, the field of Computational Argumentation has seen a remarkable expansion within the Natural Language Processing (NLP) community with the potential to have an impact on real world applications ranging from legal and healthcare to finance and (social) media. Our contributions to this field are threefold:
more » ... ld are threefold: we develop models for Argument Mining (AM), methods for mining Argumentation Frameworks (AFs) from text, and applications of the two in real world settings. We focus on one area in AM, namely Relation-based AM (RbAM), that deals with determining the argumentative relation of attack or support between texts. We bring together existing datasets available for RbAM, covering various domains and topics, and propose neural models for the cross-domain RbAM classification task. We show that our models perform homogeneously over all existing datasets for relation prediction in AM overcoming the issue of AM models that are hardly portable from one application domain to another. We also present methods for extracting AFs from text, using the developed AM models in combination with NLP techniques, and Abstract Argumentation (AA)-driven Case-Based Reasoning (CBR), an existing reasoning formalism. We apply AM in detecting deceptive reviews and in extracting arguments from a movie review website, Rotten Tomatoes, one of the most trusted websites for movies, whilst providing explanations extracted from the AFs mined from reviews. We apply AA-CBR in the Sentiment Analysis task to determine the sentiment polarity of short texts and show empirically that reasoning with AFs leads to an improvement in results compared to standard machine learning models whilst also illustrating the portability of explanations to binary classification problems.
doi:10.25560/81683 fatcat:p63ehtuws5g6tdubhgx7ah77ru