Two-Stage Ranking Using HyDE and SimCSE for Paper Retrieval
Keywords: KDD Cup, academic knowledge graph; benchmark; academic graph mining
TL;DR: OAG-AQA competition 6th place solution: Two-Stage Ranking using HyDE and SimCSE for Paper Retrieval
Abstract: The overall goal of academic data mining is to increase our understanding of the development, nature, and trends in science. Academic
data mining has the potential to extract enormous scientific, technical, and educational value. The organizers of the KDD Cup 2024 published the OAG Benchmark for Academic Graph Mining, and within it, the OAG-AQA competition focused on paper retrieval. In this paper, we present our solution that won 6th place as DOCOMOLABS in the public leaderboard in the OAG-AQA competition.
Our solution proposes a two-stage prediction model. In stage 1, we use query augmentation and contrastive learning to create candidates for paper retrieval. In stage 2, the encoder-based language model is used as a re-ranker to train binary classification and eventually ensembling the predictions of several models.
The source code of our solution is available at https://github.com/NTT-DOCOMO-RD/kddcup2024-oag-challenge-ind-7th-aqa-7th-solution-nttdocomolabs/tree/main/AQA
Submission Number: 5
Loading