Multi-View Learning: Decoding Emotions in the "Hold 不住" Era of Code-Switching

TL;DR

Deep learning for emotion detection often assumes a "pure" linguistic environment. However, real-world social media is a messy blend of languages—Code-Switching. This paper proposes a multi-view learning framework that treats Chinese, English, and a translated "Bilingual" version as distinct perspectives. By leveraging a co-training algorithm, the model learns to bridge the gap between languages, achieving a significant performance boost in identifying emotions like happiness, sadness, and surprise in mixed-language posts.

Problem & Motivation: The Bilingual Emotional Gap

As global social media users frequently mix languages (e.g., "This party was so high 翻全场"), traditional monolingual NLP models face a dilemma. If they focus only on one language, they lose the context of the other; if they treat them as a single string, the statistical sparsity of the secondary language (English in Chinese Weibo) often dilutes the signal.

The authors observed that:

• Monolingual models fail: An English-only model on Weibo data performs poorly (F1-score ~0.32) because English is often used only for specific emotional emphasis.
• Semantic Integration is Key: Emotions in code-switching are not just additive; they are often integrated into specific hybrid phrases that require a unified bilingual understanding.

Methodology: The Three-Lens Perspective

The core innovation lies in the Multi-View Framework, which decomposes a single post into three feature spaces:

1. Monolingual Views (CN & EN): Separate features extracted from the native Chinese and English text.
2. Bilingual View: This is the "bridge." Since Chinese dominates the dataset, the authors use a Statistical Machine Translation (SMT) strategy to map English words into the Chinese semantic space.
   • They don't just stop at translation; they use Sentiment Lexicons and Synonym Dictionaries to ensure that "Happy" (EN) and "开心" (CN) are mapped to the same emotional pivot.

Model Architecture

The framework uses a Co-Training algorithm. Starting with a small set of labeled data, it trains three separate classifiers ($C_{CN},C_{EN},C_{BI}$). These classifiers then "label" unlabeled data, and the most confident predictions are added back to the training set for the next iteration.

Experiments & Results

The researchers tested their approach on 4,195 manually annotated Weibo posts.

Performance Comparison

The multi-view approach consistently outperformed both supervised baselines and simpler semi-supervised methods.

Key Insights from results:

• Language Bias: Happiness occurs more frequently in English segments compared to sadness, suggesting a cultural/linguistic preference in how users switch codes.
• The Power of Synergy: Even though the English view alone was weak, it provided complementary "clues" that, when combined with the bilingual view, allowed the model to generalize better than the Chinese-only model (ME-CN).

Critical Analysis & Conclusion

Takeaway: This work proves that in multilingual social media, treating languages as "separate but equal" views is superior to treating them as a single noisy sequence. The use of SMT to create a "middle-ground" bilingual view effectively addresses the data sparsity of the secondary language.

Limitations:

• The translation is word-by-word, which may miss complex idiomatic expressions or slang.
• The framework relies on statistical ME (Maximum Entropy) models, which have since been surpassed by Transformer-based architectures (like mBERT).

Future Outlook: The methodology of "Bilingual Views" is highly applicable to modern Large Language Models (LLMs). By explicitly prompting models to look at code-switching through multiple lingual lenses, we can likely improve the emotional intelligence of AI in globalized digital spaces.