Scholarlink

Code-mixed text is often used in multi-lingual countries, where people mix more than one language in a single statement to express their thoughts freely on online media platforms. However, the existing code-mixed datasets that are available are small in size and highly skewed towards the non-sarcastic class. The imbalanced data leads to poor predictive performance such as low F1 score when they are trained on machine-learning models. Thus, in this research, we make an attempt to correctly handle imbalanced Hindi-English (Hi-Eng) code-mixed data to detect sarcasm. The proposed approach uses hybrid oversampling technique (SMOTE-ENN) during 10 fold stratified cross validation. We experimented and performed comparative analysis on our proposed approach with five machine-learning models say Multinomial Naïve Bayes, Support Vector Machine, Logistic Regression, Random Forest Tree and Extreme Gradient Boosting. The F1-score of the proposed approach was compared with the baseline approach. We observed that our proposed approach shows new insight into handling the imbalanced data.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Novel Approach to Handle Imbalanced Code-Mixed Languages Using SMOTE-ENN During Stratified Cross Validation

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Uma Ojha,
  • Ajay Kumar Yadav

摘要

Code-mixed text is often used in multi-lingual countries, where people mix more than one language in a single statement to express their thoughts freely on online media platforms. However, the existing code-mixed datasets that are available are small in size and highly skewed towards the non-sarcastic class. The imbalanced data leads to poor predictive performance such as low F1 score when they are trained on machine-learning models. Thus, in this research, we make an attempt to correctly handle imbalanced Hindi-English (Hi-Eng) code-mixed data to detect sarcasm. The proposed approach uses hybrid oversampling technique (SMOTE-ENN) during 10 fold stratified cross validation. We experimented and performed comparative analysis on our proposed approach with five machine-learning models say Multinomial Naïve Bayes, Support Vector Machine, Logistic Regression, Random Forest Tree and Extreme Gradient Boosting. The F1-score of the proposed approach was compared with the baseline approach. We observed that our proposed approach shows new insight into handling the imbalanced data.