Lets Argue Understanding And Generating Arguments

30 Nov 2020 - Arun Balajiee

Talk Speaker: Iryna Gurevych

Talk Date: 11/30/2020

Argument mining is the method of identifying and classifying sentences that make claims and their supporting statements from a text. Argument mining is similar to sentiment analysis, in that sentiment analysis is about classifying the text sentiment as positive or negative while argument mininig is about classifying statements that are pro, con or neutral in their support of a claim. Arguments have traditionally been a way to get to the truth through discussion and debate. Arguments can be simple with a direction relationship of claim and reason between two statements, or can be complicated wherein a claim is support by a reason that is supported by prior knowledge.

Dr. Gurevych, a professor in Germany and her students are identifying ways to classify states that are in support or against and argument. Further, they worked on implemeting the generation of these supporting or opposing statement claims using a transformer based neural model. The model to classify sentences uses a training set labelled by two expert annotators and using crowdsourcing to validate the dataset. Using this data-driven classification technique, they were able to identify the argument stance (pro or con), its aspects and within a set of topics. The dataset that they collected for each topic also affected the size of the dataset. Next, they used a transformer based neural model that usese these three features from statements (topic, aspect and stance) to generate statements that are for or against claims and could achieve near human accuracy. The baseline model they used was a baseling information retrieval system. They are looking towards developing a new model that can handle convincing arguments and improve the argument quality of the generated argument to be persuasive during argument-counter argument scenarios in a debate.

This talk was especially useful to look critical at the simpler aspect of natural langauge processing in which the problem is well definted and what is being pushed in the state of the art is really the accuracy of these models. In that sense, Dr. Gurevych’s work is novel and has possibilities in the space of healthcare where doctors could use the engine to generate convincing arguments to persuade interventions in patients’ lives.