Domain adaptive learning for multi realm sentiment classification on big data.

Area of Science:

• Natural Language Processing
• Machine Learning
• Computational Linguistics

Background:

• Traditional sentiment analysis models struggle with domain-specific nuances and large datasets.
• Supervised learning requires extensive, time-consuming data labeling, often leading to insufficient datasets.

Purpose of the Study:

• To evaluate the effectiveness of transfer learning in Multi-Domain Sentiment Classification (MDSC).
• To assess the impact of domain adaptation on sentiment analysis performance.
• To quantify the enhancement transfer learning brings to sentiment analysis outcomes.

Main Methods:

• Utilized transfer learning models: BERT, RoBERTa, ELECTRA, and ULMFiT.
• Employed a Multi-Domain Sentiment Classification (MDSC) technique for cross-domain learning.
• Compared transformer models against LSTM and CNN architectures.

Main Results:

• Transfer learning models demonstrated improved performance in sentiment analysis across diverse domains.
• Domain adaptation using transfer learning effectively addressed challenges in unlabeled target domains.
• Experiments on five datasets (Hotel Reviews, Movie Reviews, Tweets, CSC, BCC) confirmed model efficacy.

Conclusions:

• Transfer learning significantly enhances sentiment analysis, particularly in scenarios with limited labeled data and diverse domains.
• The MDSC technique offers a robust solution for cross-domain sentiment classification.
• Transformer-based models, enhanced by transfer learning, show superior performance in sentiment analysis tasks.