The Dark Side of the Language: Pre-trained Transformers in the DarkNet
Leonardo Ranaldi, Aria Nourbakhsh, Arianna Patrizi, Elena Sofia Ruzzetti, Dario Onorati, Francesca Fallucchi, Fabio Massimo Zanzotto
TL;DR
The paper investigates how pre-trained Transformers perform on definitely unseen DarkWeb sentences and compares them with lexical and syntactic models. It finds that Transformers underperform without domain adaptation and only reach expected high performance after extreme domain adaptation via masked language modeling on the novel corpus, suggesting pre-training exposure to many sentence patterns provides unexpected advantages. The results indicate that the benefits of large pre-training corpora may arise from exposure to many possible sentences, effectively reducing out-of-domain risk when confronted with truly unseen data. The authors advocate for transparency in pre-training data and propose future work to understand what MLM learning captures and how to quantify a model's knowledge about a given text corpus.
Abstract
Pre-trained Transformers are challenging human performances in many NLP tasks. The massive datasets used for pre-training seem to be the key to their success on existing tasks. In this paper, we explore how a range of pre-trained Natural Language Understanding models perform on definitely unseen sentences provided by classification tasks over a DarkNet corpus. Surprisingly, results show that syntactic and lexical neural networks perform on par with pre-trained Transformers even after fine-tuning. Only after what we call extreme domain adaptation, that is, retraining with the masked language model task on all the novel corpus, pre-trained Transformers reach their standard high results. This suggests that huge pre-training corpora may give Transformers unexpected help since they are exposed to many of the possible sentences.