Clustering Document Parts: Detecting and Characterizing Influence Campaigns from Documents

We propose a novel clustering pipeline to detect and characterize influence campaigns from documents. This approach clusters parts of document, detects clusters that likely reflect an influence campaign, and then identifies documents linked to an influence campaign via their association with the high-influence clusters. Our approach outperforms both the direct document-level classification and the direct document-level clustering approach in predicting if a document is part of an influence campaign. We propose various novel techniques to enhance our pipeline, including using an existing event factuality prediction system to obtain document parts, and aggregating multiple clustering experiments to improve the performance of both cluster and document classification. Classifying documents after clustering not only accurately extracts the parts of the documents that are relevant to influence campaigns, but also captures influence campaigns as a coordinated and holistic phenomenon. Our approach makes possible more fine-grained and interpretable characterizations of influence campaigns from documents.

We have presented a new approach to finding influence campaigns, which relies on four core features: (1) we cluster parts of documents; (2) we classify clusters of parts of documents using non-lexical features; (3) we relate the classification result back to documents; (4) we use cluster aggregation, the use of many clustering runs over the same dataset, to augment training data for the cluster classifier. The resulting classification of the documents does not only show a predicted label for the document (part of influence campaign or not), but it also shows which parts of the document are responsible for this classification. We believe that our general approach can profit other document classification tasks, including detecting scientific influence in published papers, or themes in literature.

There are several avenues for possible future work and we list three below. (1) Datasets. Given the increasing importance of detecting influence campaigns, we hope there will be more datasets annotated on the document collection level for an influence campaign. (2) Incorporating non-textual information. Our current pipeline is a text-only system. Leveraging non-textual information, such as social interactions and the authors' past activities, may help us create a more complicated and comprehensive system (e.g., using graph neural network) that enhances the accurate and reliable detection of influence campaigns. However, such work cannot be possible without good datasets. (3) Automatic characterization of influence campaigns. Our work captures influence campaigns by the high-influence clusters, which may contain a large number of semantically related document parts, possibly with noise. To fully make sense of these clusters, we need to have some automatic ways of characterizing them in a fine-grained and interpretable way aligned with the downstream needs. Our preliminary experiments show that LLMs may be a potential option.