An Effective BERT-Based Pipeline for Twitter Sentiment Analysis: A Case Study in Italian

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Social Media Analysis

Background:

• Growing interest in sentiment analysis for social media, particularly tweets.
• Current state-of-the-art models often require training from scratch on tweet-specific data.
• Challenges in handling Twitter jargon, including emojis and emoticons.

Purpose of the Study:

• To introduce a novel, two-step approach for Twitter sentiment analysis.
• To improve sentiment analysis by transforming tweet jargon into plain text.
• To leverage pre-trained language models on plain text for enhanced performance and language independence.

Main Methods:

• A two-step process: 1. Transforming tweet jargon (emojis, emoticons) into language-independent plain text. 2. Classifying transformed tweets using BERT pre-trained on plain text corpora.
• Utilizing language-independent procedures for jargon transformation.
• Employing BERT models pre-trained on extensive plain text datasets.

Main Results:

• Demonstrated effectiveness of the proposed approach in a case study on Italian sentiment analysis.
• Achieved competitive or superior performance compared to existing Italian solutions.
• The approach showed promise for application in other languages due to its general methodology.

Conclusions:

• The proposed method offers an effective and adaptable solution for Twitter sentiment analysis.
• Leveraging pre-trained BERT models on plain text is advantageous over training from scratch on tweet-only data.
• The approach's language-independent nature and reliance on readily available pre-trained models make it a promising tool for multilingual sentiment analysis.