[Paper Review] Analysis of Disinformation and Fake News Detection Using Fine-Tuned Large Language Model

Bohdan M. Pavlyshenko|arXiv (Cornell University)|Sep 9, 2023

Misinformation and Its Impacts13 citations

TL;DR

The paper fine-tunes Llama 2 using PEFT/LoRA to perform multitask disinformation analysis, including fake news detection, propaganda narrative highlighting, fact checking, manipulation analytics, and named-entity sentiment extraction. It demonstrates qualitative and a notable quantitative result (F1=0.95) on Twitter data for fake news detection and narrative analysis.

ABSTRACT

The paper considers the possibility of fine-tuning Llama 2 large language model (LLM) for the disinformation analysis and fake news detection. For fine-tuning, the PEFT/LoRA based approach was used. In the study, the model was fine-tuned for the following tasks: analysing a text on revealing disinformation and propaganda narratives, fact checking, fake news detection, manipulation analytics, extracting named entities with their sentiments. The obtained results show that the fine-tuned Llama 2 model can perform a deep analysis of texts and reveal complex styles and narratives. Extracted sentiments for named entities can be considered as predictive features in supervised machine learning models.

Motivation & Objective

• Motivate the need to detect disinformation and fake news in social media and news flows.
• Assess whether a fine-tuned Llama 2 LLM can multitask in disinformation analysis tasks.
• Develop a PEFT/LoRA-based fine-tuning pipeline enabling efficient training on small datasets.

Proposed method

• Fine-tune Llama 2-7B-chat using PEFT/LoRA with 4-bit quantization and a trainer like SFTTrainer.
• Use prompt instructions to guide the LLM to analyze text for disinformation, highlight main points, summarize, and extract named entities with sentiments.
• Incorporate training data from fake news datasets and Russian propaganda narratives; split data into training and validation (25% for validation).
• Evaluate outputs on tasks including disinformation analysis, propaganda narrative highlighting, fact checking, fake news detection, manipulation analytics, and entity sentiment extraction.

Figure 1: Time series of tweets for the query ’ukraine’.

Experimental results

Research questions

• RQ1 Can a fine-tuned Llama 2 model perform multitask disinformation analysis with a structured, potentially JSON-formatted output?
• RQ2 Do PEFT/LoRA-based fine-tuning and 4-bit quantization enable effective, resource-efficient adaptation for disinformation tasks?
• RQ3 Are extracted entity sentiments useful as predictive features in downstream models for media or market analysis?
• RQ4 How well does the model identify and analyze propaganda narratives and manipulation in text?
• RQ5 What are the limitations and accuracy concerns of the fine-tuned model for disinformation analytics?

Key findings

• A fine-tuned Llama 2 model can perform multitask text analysis with structured outputs.
• The approach enables extraction of named entities with sentiments suitable for downstream models.
• PAEFT/LoRA with 4-bit quantization enables cost-effective fine-tuning on small datasets.
• The model demonstrated narrative analysis and fact-checking capabilities on excerpts regarding Ukraine and related topics.
• A qualitative assessment indicates some inaccuracies in outputs, suggesting improvements via better datasets and RLHF.

Figure 2: Time series of tweets for the thematic field ’ukraine nazi’.