Papers by Eugene Jang
Shedding New Light on the Language of the Dark Web (2022.naacl-main)
Copied to clipboard
| Challenge: | Existing studies on text classification of the Dark Web have been ineffective due to its inherent characteristics. |
| Approach: | They propose a publicly available Dark Web dataset tailored towards text-based analysis. |
| Outcome: | The proposed method compares with an existing public Dark Web dataset and evaluates its suitability for various use cases. |
Improbable Bigrams Expose Vulnerabilities of Incomplete Tokens in Byte-Level Tokenizers (2025.emnlp-main)
Copied to clipboard
| Challenge: | Recent studies have revealed that tokenizers can be exploited to elicit unwanted behavior. |
| Approach: | They propose to exploit incomplete tokens with stray bytes to exploit their dependency . they propose to use improbable bigrams to exploit the dependency of their adjacent tokens . |
| Outcome: | The proposed tokenizers can be exploited to elicit unwanted behavior in language models. |
Ignore Me But Don’t Replace Me: Utilizing Non-Linguistic Elements for Pretraining on the Cybersecurity Domain (2024.findings-naacl)
Copied to clipboard
| Challenge: | Existing methods to train domain expertise for cybersecurity text domains are expensive to train and run. |
| Approach: | They propose to use pretraining methods to account for non-linguistic elements in cybersecurity texts and evaluate their effectiveness through downstream tasks and probing tasks. |
| Outcome: | The proposed strategy outperforms the commonly taken approach of replacing NLEs and outperformed other cybersecurity PLMs on most tasks. |
Generating Negative Samples by Manipulating Golden Responses for Unsupervised Learning of a Response Evaluation Model (2021.naacl-main)
Copied to clipboard
| Challenge: | Existing metrics that rely on comparisons to a set of known correct responses do not account for the variety of responses and therefore correlate poorly with human judgment. |
| Approach: | They propose a method of manipulating a golden response to create a new negative response that is designed to be inappropriate within the context while maintaining high similarity with the original golden response. |
| Outcome: | The proposed model can be made using unsupervised learning for the next-utterance prediction task on English datasets and shows that using the negative samples alongside random negative samples can increase the model’s correlation with human evaluations. |
WinoQueer: A Community-in-the-Loop Benchmark for Anti-LGBTQ+ Bias in Large Language Models (2023.acl-long)
Copied to clipboard
| Challenge: | Existing benchmarks for measuring anti-LGBTQ+ bias are poorly defined and insufficiently grounded in real-world harms. |
| Approach: | They propose a bias benchmark that is community-sourced and generates a community survey. |
| Outcome: | The proposed method is community-sourced and improves on WinoQueer-v0. |
DarkBERT: A Language Model for the Dark Side of the Internet (2023.acl-long)
Copied to clipboard
| Challenge: | Recent research suggests that there are clear differences in the language used in the Dark Web compared to that of the Surface Web. |
| Approach: | They propose a language model that is pretrained on Dark Web data to combat extreme lexical diversity. |
| Outcome: | The proposed model outperforms existing models and may be useful for future research on the Dark Web. |