Papers by David Gros
Cross-Domain Detection of GPT-2-Generated Technical Text (2022.naacl-main)
Copied to clipboard
| Challenge: | Recent advances in techniques for generating realistic synthetic content pose a diverse set of problems with significant societal consequences. |
| Approach: | They propose to use paragraph-level detectors to detect tampering of full-length documents under a variety of threat models to detect machine-generated text. |
| Outcome: | The proposed detectors can detect the tampering of full-length documents under a variety of threat models. |
Robots-Dont-Cry: Understanding Falsely Anthropomorphic Utterances in Dialog Systems (2022.emnlp-main)
Copied to clipboard
| Challenge: | Dialog systems often output human-like responses, but some are impossible for a machine to say. |
| Approach: | They collect ratings on the feasibility of 900 two-turn dialogs from 9 data sources . they build classifiers and explore how modeling configuration might affect output permissibly . |
| Outcome: | The proposed model can be used to train human-like dialogs, but it is not anthropomorphic. |
The R-U-A-Robot Dataset: Helping Avoid Chatbot Deception by Detecting User Questions About Human or Non-Human Identity (2021.acl-long)
Copied to clipboard
| Challenge: | We analyze 2,500 phrasings related to the intent of “Are you a robot?” and 2,500 adversarially selected utterances to determine whether systems are non-human. |
| Approach: | They analyze 2,500 phrasings related to the intent of "Are you a robot?" and 2,500 adversarially selected utterances to determine whether systems are non-human. |
| Outcome: | The proposed model and two systems fail to confirm non-human intent, and the proposed model is complex. |