Machine Learning and Lexical Rule-Based Cost-Efficient Emotion Annotation of Hinglish Utterances

Area of Science:

• Computational Linguistics
• Natural Language Processing
• Affective Computing

Background:

• Emotion annotation in code-mixed languages like Hinglish faces challenges due to linguistic complexity and limited resources.
• Existing methods often struggle with the nuances of Hinglish, impacting accuracy and cost-efficiency.

Purpose of the Study:

• To develop a cost-efficient, high-accuracy emotion annotation framework for Hinglish text.
• To integrate lexical rules, machine learning, and active learning for improved annotation.
• To address the unique challenges of emotion analysis in code-mixed languages.

Main Methods:

• A hybrid active learning framework combining lexical rules, machine learning, and iterative expert feedback.
• Utilized bilingual emotion dictionaries, subword tokenization for compound terms, and active learning for sample prioritization.
• Grounded in Discrete Emotions Theory and Cognitive Appraisal Theory.

Main Results:

• Achieved 81% accuracy (F-score: 0.76) on a 19,000 Hinglish tweets dataset related to war and conflict.
• Reduced operational costs by 40% compared to traditional manual annotation.
• Lexical rules resolved 89% of code-switching ambiguities, with iterative refinements increasing accuracy from 72% to 81%.

Conclusions:

• The hybrid framework significantly enhances emotion annotation accuracy in Hinglish while reducing costs.
• Integrating rule-based methods with active learning and machine learning is effective for code-mixed language processing.
• Automated preprocessing of emojis, hashtags, and slang contributes to system efficiency.