Sentiment Analysis on the People's Daily

Jiwei Li, Eduard H. HovyEMNLP 2014 · OpenReview

Abstract

We propose a semi-supervised bootstrapping algorithm for analyzing China’s foreign relations from the People’s Daily. Our approach addresses sentiment target clustering, subjective lexicons extraction and sentiment prediction in a unified framework. Different from existing algorithms in the literature, time information is considered in our algorithm through a hierarchical bayesian model to guide the bootstrapping approach. We are hopeful that our approach can facilitate quantitative political analysis conducted by social scientists and politicians.

Introduction

Analyzing political sentiment in official state media like the People’s Daily—the primary mouthpiece of the CCP—offers invaluable insights for sociologists and historians. However, this task presents unique challenges for standard Natural Language Processing (NLP). Political discourse is heavily laden with metaphors and rhetoric (e.g., "paper tiger"), where standard sentiment classifiers often fail. Furthermore, single articles frequently contain multiple targets with opposing sentiments, such as praising an ally while condemning an "imperialist" foe. Finally, political language is dynamic; "enemies" of the 1950s can evolve into "partners" by the 2000s.

To address these challenges, we developed a semi-supervised bootstrapping algorithm that tracks 60 years of China’s foreign relations. By incorporating a Hierarchical Bayesian model that accounts for temporal shifts in political tone, the system can decode complex metaphors and accurately distinguish between targets even within the same document.

Methodology

The proposed framework utilizes a semi-supervised bootstrapping approach requiring only a minimal "seed set" of subjective words. It comprises three main components:

  1. Target and Expression Extraction: We employ semi-Markov Conditional Random Fields (semi-CRFs) to identify multi-word entities (e.g., "Vietnamese People’s Army") more effectively than standard word-level CRFs.
  2. Hierarchical Bayesian Markov Model: This core innovation assumes sentiment coherence over time. Using a Poisson process, it models sentiment scores across four levels: time-period, document, sentence, and term. This allows the system to leverage the known sentiment of a specific era to resolve the meaning of unknown words.
  3. Bootstrapping with Error Prevention: The system iteratively expands its sentiment lexicon and target entities. To prevent "semantic drift," a strict selection policy is applied, bootstrapping only from sentences with exactly one target and one expression.

The model was trained and evaluated on a 60-year archive of the People’s Daily (1950–2010), supplemented by the Tsinghua Foreign Relations Reports—a gold-standard dataset of monthly quantitative reports on China's relations with seven key nations.

Results and Discussion

The model’s performance was evaluated against the Tsinghua gold standard using Pearson Correlation. Our proposed model achieved a correlation of 0.895, significantly outperforming baselines such as SVR (0.482) and supervised LDA (0.427).

Sentiment scores over time for six major countries, showing historical shifts in diplomatic relations.

A key finding is the importance of temporal context. A "No-time" version of the algorithm scored only 0.808, demonstrating that modeling temporal coherence is vital for political analysis. The model successfully mapped major historical shifts, such as the improvement in Sino-American relations during the 1970s coinciding with the decline in Sino-Soviet relations.

Top coreference targets for the U.S. and Soviet Union/Russia across decades, illustrating the evolution of political rhetoric.

The model also excels at sentiment-based clustering. Instead of simple coreference resolution, it groups entities by the attitude expressed toward them. For instance, "US imperialism" and "Johnson Government" are clustered together due to the consistent negative sentiment directed at them during specific periods.

Conclusion

While the model handles metaphors better than previous methods, identifying deeply nested rhetoric remains a challenge. The current strict sentence selection also sacrifices data volume for accuracy. Future work will focus on processing more complex sentence structures and extending the analysis to non-state media where sentiment may be more fragmented.

Summary updated

Abstract

We propose a semi-supervised bootstrapping algorithm for analyzing China’s foreign relations from the People’s Daily. Our approach addresses sentiment target clustering, subjective lexicons extraction and sentiment prediction in a unified framework. Different from existing algorithms in the literature, time information is considered in our algorithm through a hierarchical bayesian model to guide the bootstrapping approach. We are hopeful that our approach can facilitate quantitative political analysis conducted by social scientists and politicians.
BibTeX
@inproceedings{li2014sentiment,
  author = {Jiwei Li and Eduard H. Hovy},
  title = {{Sentiment Analysis on the People's Daily}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2014},
  url = {https://openreview.net/forum?id=BkVU7zMd-r},
}
Figure