Scholarlink

This paper addresses the challenge of automated citation prediction within the SCIDOCA 2025 Shared Task, focusing on Citation Discovery (Task 1) and Citation Sentence Prediction (Task 3). We propose an information retrieval approach utilizing transformer models to assess semantic similarity between query passages/sentences and candidate citations, which are represented by the concatenation of their titles and abstracts. Experiments were conducted using bge-large-en-v1.5, gte-qwen2-1.5b-instruct, and bge-reranker-large. Results demonstrate that bge-large-en-v1.5 outperforms gte-qwen2-1.5b-instruct in F1 score for Task 1, and fine-tuning bge-large-en-v1.5 achieved F1 score at 45%. Notably, bge-reranker-large significantly outperforms bge-large-en-v1.5 in Task 3, achieving F1 score of 66%, highlighting the effectiveness of cross-encoder architectures for sentence-level citation prediction.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Embedding-Based Retrieval Approaches for Automated Citation Prediction

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Dat Le,
  • Son Nguyen

摘要

This paper addresses the challenge of automated citation prediction within the SCIDOCA 2025 Shared Task, focusing on Citation Discovery (Task 1) and Citation Sentence Prediction (Task 3). We propose an information retrieval approach utilizing transformer models to assess semantic similarity between query passages/sentences and candidate citations, which are represented by the concatenation of their titles and abstracts. Experiments were conducted using bge-large-en-v1.5, gte-qwen2-1.5b-instruct, and bge-reranker-large. Results demonstrate that bge-large-en-v1.5 outperforms gte-qwen2-1.5b-instruct in F1 score for Task 1, and fine-tuning bge-large-en-v1.5 achieved F1 score at 45%. Notably, bge-reranker-large significantly outperforms bge-large-en-v1.5 in Task 3, achieving F1 score of 66%, highlighting the effectiveness of cross-encoder architectures for sentence-level citation prediction.