Abstract
<title>Abstract</title> <p>In the era of artificial intelligence, algorithms have become a core driver of scientific research and innovation. Identifying algorithm entities and their semantic relations is important for understanding the development and innovation trajectories of algorithms. Focusing on natural language processing (NLP), this study identifies semantic relations among algorithm entities in full-text papers and analyzes their interaction patterns and evolution. We first collected full-text papers from three major NLP conferences and constructed an updated corpus. Pre-trained models were then fine-tuned to identify algorithm entities and sentences. After that, we defined an algorithm semantic relation framework and manually annotated a gold-standard corpus. Various large language models (LLMs) and pre-trained language models (PLMs) were used for identifying sentence-level relations among algorithms. Finally, we analyzed representative interacting algorithms and the evolution trends of algorithm relation networks. The results show that SciBERT achieves the best overall performance in relation identification, while LLMs with few-shot learning and chain-of-thought reasoning also obtain competitive results. In the NLP domain, algorithm semantic relations are dominated by the Compare relation, and multiple potential interaction motivations underlie the same kinds of relation. All of the relation-specific networks show expanding scale, deepening interaction, and accelerated renewal over time, reflecting the increasingly intense ecosystem of algorithm competition in the NLP field. This study provides data resources and methodological references for relation identification and graph construction among knowledge entities.</p>