Papers by Kent Chang
Speak, Memory: An Archaeology of Books Known to ChatGPT/GPT-4 (2023.emnlp-main)
Copied to clipboard
| Challenge: | a recent study has shown that open AI models memorize a wide collection of copyrighted materials . however, these models also present a challenge for establishing the validity of results . |
| Approach: | They propose to use a name cloze membership inference query to infer books that are known to ChatGPT and GPT-4. |
| Outcome: | The proposed model performs better on memorized books than on non-memorized books for downstream tasks. |
Dramatic Conversation Disentanglement (2023.findings-acl)
Copied to clipboard
| Challenge: | a new dataset is available for studying conversation disentanglement in movies and TV series . a recent study focused on IRC chatroom dialogues, but movies and television show provide a space for study . |
| Approach: | They propose a dataset for studying conversation disentanglement in movies and TV series . they operationalize a conversational thread and apply the best-performing model to 808 movies . |
| Outcome: | The proposed model disentangles 808 movies from 10,033 dialogue turns . the best-performing model is compared with previous models . |