Abstract
A large number of conflict events are affecting the world all the time. In order to analyse such conflict events effectively, this paper presents a Classification-Aware Neural Topic Model (CANTM-IA) for Conflict Information Classification and Topic Discovery. The model provides a reliable interpretation of classification results and discovered topics by introducing interpretability analysis. At the same time, interpretation is introduced into the model architecture to improve the classification performance of the model and to allow interpretation to focus further on the details of the data. Finally, the model architecture is optimised to reduce the complexity of the model.
Introduction
In the high-stakes domain of international conflict monitoring, speed and accuracy are paramount. Organizations such as the International Food Policy Research Institute (IFPRI) must classify thousands of reports to respond to crises effectively. While deep learning models like BERT offer high accuracy, they often operate as "black boxes," providing little insight into the specific textual evidence driving a classification. This lack of transparency can hinder trust and actionability. Furthermore, standard topic models are frequently "polluted" by neutral words (e.g., "report," "city") that do not help distinguish between conflict types.
Methodology
The proposed CANTM-IA (Classification-Aware Neural Topic Model with Interpretable Analysis) addresses these challenges by integrating interpretability directly into the model's learning process.
The model introduces three key innovations:
- Saliency-Based Rationales: By analyzing attention weights in the final BERT layer, the model identifies "rationales"—the most influential words for a given classification task.
- Rationale-Guided Reconstruction: Unlike standard Variational Auto-Encoders (VAEs) that attempt to reconstruct the entire bag-of-words, CANTM-IA focuses on reconstructing only high-saliency rationales. This acts as a sophisticated noise filter, focusing topic discovery on the most meaningful content.
- Architectural Optimization: The researchers identified and removed redundant structures from the original CANTM design, specifically regarding how classification probabilities were fed back into the decoder, resulting in a more efficient model.
Experimental Setup
The model was evaluated using the ACLED (Armed Conflict Location & Event Data Project) dataset, a professionally curated collection of conflict events. The study utilized data from 2019 to 2022, covering six primary categories: Battles, Explosions/Remote Violence, Protests, Riots, Strategic Developments, and Violence Against Civilians. To ensure robustness, the WMT News Crawl dataset was included as "out-of-domain" data. The final dataset comprised over 522,000 items.
Results and Discussion
CANTM-IA outperformed both the standalone BERT baseline and the original CANTM model, achieving an accuracy of 0.9780 and an F1 score of 0.9791. Beyond numerical improvements, the model demonstrated significant qualitative superiority in interpretability.
As illustrated in Figure 6, the rationales extracted by CANTM-IA are more precise and focused on conflict-related information (e.g., identifying "soldiers," "destroyed," and "church") compared to baselines that often highlight irrelevant function words like "as" or "reported." This focus results in "cleaner" topic clouds, where discovered topics consist of sharp, category-specific terms (e.g., "clashed," "fatalities," "airstrikes") rather than generic vocabulary.
Conclusion
CANTM-IA demonstrates that interpretability is not merely a post-hoc diagnostic tool but a functional component that can enhance model performance by filtering linguistic noise. Future work will involve adapting the model to noisier, less structured datasets and exploring its application in related NLP tasks such as stance detection and rumor verification.
Summary updated