NLP_SEMINAR Telegram 131
Sometimes you just want to participate in scientific ML competitions and publish papers with your results. That's what we did.

We would like to present you our papers:
- AIpom at SemEval-2024 Task 8: Detecting AI-produced Outputs in M4 (SemEval2024, NAACL workshop)
- Papilusion at DAGPap24: Paper or Illusion? Detecting AI-generated Scientific Papers (SDProc, ACL workshop)

🔸 SemEval-2024 task 8 required to find the position of the first machine-generated word in given texts. The test dataset also included texts from a different domain to check the solutions robustness. We proposed a novel method for this problem and called it AIpom. Our approach was based on a pipeline of decoder-only and encoder-only models utilized sequentially. We generated preliminary predictions using the LLM and passed them to the encoder model.

🔸 DAGPap24 task was to classify scientific texts tokens into 4 categories: (i) human-written, (ii) modified through an NLTK-based synonym replacement, (iii) produced by ChatGPT, and (iv) summarized. We fine-tuned encoder models to predict the corresponding labels, i.e. each model was trained on a sequence labeling task, predicting labels for every token in the input sequence. We also conduct ablation studies to analyze the effect of the detector configurations on the performance.

Results:
SemEval2024 - 2nd place on the subtask C leaderboard
DAGPap24 - 6th place on the leaderboard

Paper links:
- AIpom 🎆
- Papilusion



tgoop.com/nlp_seminar/131
Create:
Last Update:

Sometimes you just want to participate in scientific ML competitions and publish papers with your results. That's what we did.

We would like to present you our papers:
- AIpom at SemEval-2024 Task 8: Detecting AI-produced Outputs in M4 (SemEval2024, NAACL workshop)
- Papilusion at DAGPap24: Paper or Illusion? Detecting AI-generated Scientific Papers (SDProc, ACL workshop)

🔸 SemEval-2024 task 8 required to find the position of the first machine-generated word in given texts. The test dataset also included texts from a different domain to check the solutions robustness. We proposed a novel method for this problem and called it AIpom. Our approach was based on a pipeline of decoder-only and encoder-only models utilized sequentially. We generated preliminary predictions using the LLM and passed them to the encoder model.

🔸 DAGPap24 task was to classify scientific texts tokens into 4 categories: (i) human-written, (ii) modified through an NLTK-based synonym replacement, (iii) produced by ChatGPT, and (iv) summarized. We fine-tuned encoder models to predict the corresponding labels, i.e. each model was trained on a sequence labeling task, predicting labels for every token in the input sequence. We also conduct ablation studies to analyze the effect of the detector configurations on the performance.

Results:
SemEval2024 - 2nd place on the subtask C leaderboard
DAGPap24 - 6th place on the leaderboard

Paper links:
- AIpom 🎆
- Papilusion

BY исследовано




Share with your friend now:
tgoop.com/nlp_seminar/131

View MORE
Open in Telegram


Telegram News

Date: |

How to create a business channel on Telegram? (Tutorial) With the “Bear Market Screaming Therapy Group,” we’ve now transcended language. How to build a private or public channel on Telegram? Step-by-step tutorial on desktop: Those being doxxed include outgoing Chief Executive Carrie Lam Cheng Yuet-ngor, Chung and police assistant commissioner Joe Chan Tung, who heads police's cyber security and technology crime bureau.
from us


Telegram исследовано
FROM American