Argument Mining with LLaMA 8B
Abstract: An end-to-end argument mining (AM) pipeline takes a text as input and provides the argumentative structure of this text as output, by identifying and classifying the argument units and relations within it. In this work, we focus on LLM fine-tuning approach to AM. We model the three sub-tasks of the AM pipeline as text generation tasks. We fine-tune classical and quantized versions of LLaMA--3, the most capable open-source model available, on the benchmark Persuasive Essays (PE) dataset. We consider various contextual and structural fine-tuning modalities, where the AM sub-tasks are modeled either at the paragraph or at the essay level, with or without inclusion of additional markup tags. We achieve state-of-the-art results on all three sub-tasks, with significant improvements over previous benchmarks.
Paper Type: Short
Research Area: Sentiment Analysis, Stylistic Analysis, and Argument Mining
Research Area Keywords: argument mining, stance detection, generative models, fine-tuning, multi-task learning
Contribution Types: Model analysis & interpretability, NLP engineering experiment
Languages Studied: English
Submission Number: 2725
Loading