Natural Language Processing Applied to Forensics Information Extraction With Transformers and Graph Visualization
Fillipe Barros Rodrigues, William Ferreira Giozza, Robson de Oliveira Albuquerque, Luis Javier García Villalba
Abstract
Digital forensics analysis is a slow process mainly due to the large amount and variety of data. Some forensic tools help categorize files by type and allow automatization of tasks, like named entity recognition (NER). NER is a key component in many natural language processing (NLP) applications, such as relation extraction (RE) and information retrieval. The introduction of neural networks and transformer architectures in the last few years made it possible to develop more accurate models in different languages. This work proposes a reproducible setup to build a forensic pipeline for information extraction using NLP of texts. Our results show that it is possible to develop both NER and RE models in any language and tune its hyper-parameters to achieve state-of-art performance and build comprehensive knowledge graphs, decreasing the amount of time required for human supervision and review. We also find that solving this task in phases can further improve the performance, not only for digital investigation applications, but also for general-purpose information extraction and analysis.