Abstract:
Scientific literature aims to formally disseminate research findings, with argumentation structure playing a key role in information retrieval. However, the complexity of document structures and the scarcity of annotated datasets pose significant challenges. Additionally, there is limited research on end-to-end Argument Mining for full-text scientific papers. To address this, a web-based application will be developed to perform argumentation mining on such papers.
The research proposes a web-based application using a generative end-to-end Argument Mining (AM) model for full-text scientific texts. It incorporates the Augmented Natural Language (ANL) and argument zoning to improve Argumentative Discourse Unit (ADU) detection and relationship identification. The methodology includes data collection, feature engineering, and T5 and BART-based model training, with evaluation conducted on the Sci-Arg dataset using the Macro-F1 score.
Several models were trained from both MTL and Seq2Seq paradigms on top of the modified SciArg dataset and the Seq2Seq models outperformed the MTL model and the T5-base model in particular recorded ROUGEL score of 0.966 and macro-f1 score of 0.8192. From the benchmarking against previous studies, the T5-base model showed significant Argument Mining improvement, suggesting the strong potential for end-to-end argument mining in scientific texts.
