Work Research Skills Contact

Research Project

Multi-Pipeline Sentiment Analysis

West-African Pidgin language processing and analysis with advanced machine learning techniques.

Improving accuracy and relevance of sentiment analysis for underrepresented languages
Explore
Multi-Pipeline Sentiment Analysis

Project Overview

This research introduces a multi-pipeline approach to Sentiment Analysis, focusing on improving the accuracy and relevance of results for West-African pidgin language. Existing approaches to sentiment analysis of West-African pidgin have been fragmented, often training a new model with pidgin data, and focusing on general sentiment polarity.

Python NLP Machine Learning XLM-R AfriBerta Logistic Regression

Key Achievements

  • • Developed a holistic multi-pipeline system for exhaustive polarity analysis
  • • Achieved F1-score of 74.5% with AfriBerta model (average over 5 runs)
  • • Subject classifier achieved 81% accuracy in identifying relevant text
  • • Integrated cross-lingual model with Roberta (XLM-R) using transfer learning

Research Methodology

Multi-Pipeline Architecture

The system employs a two-stage approach: first, a subject classifier determines if the text is relevant to the target subject matter using Logistic Regression. Then, for relevant texts, a fine-tuned XLM-R model performs sentiment analysis. This approach significantly improves both accuracy and computational efficiency by filtering out irrelevant content before detailed analysis.

Data Collection and Processing

Twitter data was collected and processed into tokens for training and evaluation. The dataset was carefully curated to represent the linguistic diversity of West-African pidgin, ensuring the model's robustness across different regional variations and usage patterns.

Model Development

The sentiment analysis component uses a cross-lingual model with Roberta (XLM-R) that was fine-tuned and expanded through transfer learning in the AfriBerta model. This approach leverages pre-trained multilingual representations while adapting to the specific characteristics of West-African pidgin.

Authors

Primary Authors

  • Segun Aina
  • Joshua Etim
  • Seun Ayeni
  • Aderonke Lawal
  • Oluwatoyin Odukoya

Affiliation

Omolara Ogungbe
Obafemi Awolowo University
Department of Computer Science

Explore More Research

Discover more of my research projects exploring cutting-edge technologies and methodologies in computer science and software engineering.

← Back to Research Projects

Let's Build the Future Together

Have a project in mind?