CapstoneDesign(CSI4101) 2020 Spring Semester

"Effective Ways to Select Dataset from Large Corpus"

Team: 'input.txt' (E-mail: [email protected])

1. Joochan Kim(Leader, Code, Idea)
2. Dobreva Iva(Code)
3. Yujin Kim(Documentation, presentation)

Video & PPT

---

Idea originated from Prof. Jinyeong Yeo @ Convei lab in Yonsei Univ.

Assistance was given by Gayeon Lee @ Convei lab in Yonsei Univ.

Most of the NLP datasets are so big that lead developers spend a lot of time and costs to train model. To reduce this burden, We propose a new approach that reduces size of data to lessen time and costs needed and improves performances.

---

Model

- CEDR: Contextualized Embeddings for Document Ranking

Download

- TIM_PLUS: Two-phase Influence Maximization

Download

Dataset

- Robust04: TREC Robust document collection for Retrieval task

Download

---

How to Run?

1. Download Models and Dataset and unzip
2. run graph/graph-generator.py to make graph (Change data location at line 105 to /filename.pkl. Check README.md)
3. run TIM_PLUS using step 2's result (check README)
4. Use step 3's result to make seed.txt (Just copy the result and write into it)
5. run /graph/create-set.py (data, pkl and seed.txt needed, check README.md)
6. run /Robust-Ranker-Master/main.py (Check README)
7. Compare MAP! ^-^

---

Result

Number of dataset: 110000 -> 50000

Train time spend: 11h -> 6h