Who Blames Whom in a Crisis? Detecting Blame Ties from News Articles Using Neural Networks

Shuailong Liang, Olivia Nicol, Yue Zhang

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 the large unsupervised model 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

Following major societal disruptions—such as the 2008 financial collapse or natural disasters—society engages in complex "blame games." These narratives are not merely gossip; they actively influence regulatory outcomes, shift social hierarchies, and redefine cultural norms. For social scientists, tracking these "blame ties" (identifying who is blaming whom) is a Herculean task, traditionally requiring the manual reading of thousands of articles.

Automating this process is challenging due to context dependency (e.g., "lenient loans" only implies blame during a housing crisis), linguistic complexity (metaphors and irony), and the fact that blame is often established across multiple sentences at a document level. This paper presents a novel neural network framework to automatically extract these ties from news articles.

Methods

The researchers proposed three distinct neural architectures for the Blame Tie Extraction task:

  1. Entity Prior Model: A fully connected network (FCN) that learns the "aggressiveness" of entities, capturing the likelihood of one entity blaming another based on historical patterns.
  2. Context Model: A Bi-directional Long Short-Term Memory (BiLSTM) network. To ensure the model learns linguistic patterns of blame rather than just entity associations, entity names are masked with a special <ENT> token.
  3. Combined Model: This architecture integrates both entity embeddings and context vectors, leveraging both prior knowledge and situational text.
The BiLSTM architecture for context-aware blame tie extraction, featuring entity masking.

The models utilize GloVe and ELMo embeddings to capture deep semantic meanings. For document-level extraction, Max Pooling is used to aggregate information from multiple mentions of the same entity across an article.

Dataset and Task Definition

The study introduced a specialized dataset focused on the U.S. Financial Crisis (2007–2010), sourced from The New York Times, The Wall Street Journal, and USA Today. The dataset consists of 998 articles and 8,562 samples containing over 2,000 validated blame ties.

Examples of extracted blame ties showing the source, target, and associated causality link.

The dataset demonstrates high inter-annotator reliability with a Fleiss’s kappa of 0.8744. Additionally, a small generalization set of 13 articles from 2018 (concerning Trump, Russia, and North Korea) was used to test the model's performance on entirely new events.

Results and Insights

The best-performing Combined Model achieved an F1 score of 69.92% on the full test set. For entities already seen in the training data, performance reached 81.75% F1.

Key findings include:

  • The Power of Context: Switching to ELMo (contextualized embeddings) provided a performance boost of over 10% in F1 score compared to static embeddings, highlighting that the nuance of "blame" is heavily tied to specific word sequences.
  • Generalization: In the 2018 case study, the Context Model achieved a 72% F1 score, proving that the linguistic "fingerprints" of blame are consistent across different eras and topics.
  • Social Clustering: t-SNE visualizations of entity embeddings revealed that the model naturally clustered entities with similar political backgrounds because they tended to be blamed by the same sources.

Conclusion

While the model successfully automates a significant portion of blame tie extraction, it still faces challenges with highly implicit blame or subtle irony requiring deep world knowledge. Future work involves applying this framework to other domains, such as natural disasters or international conflicts, to determine if the "grammar of blame" remains universal. This tool represents a significant step forward in enabling social scientists to map complex narratives more efficiently.

Code and models are available at: https://github.com/Shuailong/BlamePipeline

Summary updated

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 the large unsupervised model 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.
BibTeX
@inproceedings{liang2019who,
  author = {Shuailong Liang and Olivia Nicol and Yue Zhang},
  title = {{Who Blames Whom in a Crisis? Detecting Blame Ties from News Articles Using Neural Networks}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2019},
  url = {https://openreview.net/forum?id=Ir3WR0jjDm},
}
Figure