Abstract
Blame games tend to follow major disruptions, be they financial crises, natural disasters or terrorist attacks. To study how the blame game evolves and shapes the dominant crisis narratives is of great significance, as sense-making processes can affect regulatory outcomes, social hierarchies, and cultural norms. However, it takes tremendous time and efforts for social scientists to manually examine each relevant news article and extract the blame ties (A blames B). In this study, we define a new task, Blame Tie Extraction, and construct a new dataset related to the United States financial crisis (2007–2010) from The New York Times, The Wall Street Journal and USA Today. We build a Bi-directional Long Short-Term Memory (BiLSTM) network for contexts where the entities appear in and it learns to automatically extract such blame ties at the document level. Leveraging large unsupervised models such as GloVe and ELMo, our best model achieves an F1 score of 70% on the test set for blame tie extraction, making it a useful tool for social scientists to extract blame ties more efficiently.
Introduction: The Dynamics of Crisis Narratives
Following major societal disruptions, such as the 2008 financial crisis, the ensuing "blame games" play a pivotal role in shaping future legislation and cultural norms. For social scientists, tracking these narratives involves identifying "blame ties"—structured tuples representing who is accusing whom. However, manual extraction from thousands of news articles is prohibitively slow. This paper addresses the challenge by formalizing the "Blame Tie Extraction" task and proposing a neural network-based solution.
Methodology: A Hybrid Neural Approach
The proposed architecture combines linguistic context with social knowledge to identify blame. Identifying blame is difficult because it often relies on complex syntax or domain-specific metaphors.
Contextual Modeling with BiLSTM
The core of the system is a Bi-directional Long Short-Term Memory (BiLSTM) network. To ensure the model learns generalizable patterns of accusation rather than specific entity names, the authors replace entity mentions with a generic <ENT> token. This forces the network to focus on the surrounding linguistic context. Max pooling is then used to aggregate these context vectors into a final representation.
Entity Priors and Pretrained Embeddings
Beyond linguistic context, the model incorporates an "Entity Prior Model." This feedforward network learns embeddings for specific actors, capturing their historical "aggressiveness" or typical targets. Furthermore, the researchers integrated ELMo (Embeddings from Language Models), which provided a significant boost in performance by offering deep, character-based contextual word representations.
Dataset: The U.S. Financial Crisis Benchmark
The authors introduced a manually annotated dataset focused on the 2007–2010 financial crisis. Sourced from The New York Times, The Wall Street Journal, and USA Today, the dataset includes 998 articles and 8,562 samples. The high quality of the data is evidenced by a Fleiss’s kappa of 0.8744, indicating strong inter-annotator agreement.
Results and Discussion
The combined model, integrating both context and entity priors, achieved a peak F1 score of 70%. Key insights include:
- Generalization: The model successfully identified blame ties in 2018 news (Trump/Russia/North Korea) despite being trained on 2008 data, suggesting that the linguistic "fingerprints" of blame are stable over time.
- Social Knowledge: The Entity Prior Model acts as a "social cheat sheet," significantly improving accuracy for well-known political actors.
- Semantic Clustering: t-SNE visualizations showed that the model naturally groups entities with similar political roles or blame profiles in the embedding space.
Conclusion and Future Work
This research provides a powerful tool for automating the analysis of crisis narratives. While the model excels at identifying standard blame ties, future work is needed to address high-level rhetorical devices like irony and metaphor. Expanding the dataset to include diverse crisis types, such as natural disasters, will further enhance the model's universality.
Reference: Liang, S., Nicol, O., & Zhang, Y. (2019). Who Blames Whom in a Crisis? Detecting Blame Ties from News Articles Using Neural Networks. Proceedings of the Thirty-Third AAAI Conference on Artificial Intelligence (AAAI-19).
Summary updated