「AIを応用した英語文献自動探索システムの開発とDNAバンク業務への実装」三輪 佳宏（理化学研究所 バイオリソース研究センター ）

Value of the Information & Strength of the Argument

The presentation provides valuable insights into the practical application of AI for literature screening in a biological resource center. The author clearly explains the process from data annotation to model development and implementation, highlighting challenges such as token limits and the need for careful annotation guidelines. The argumentation is based on personal experience and specific examples, making it credible. However, the lack of quantitative performance metrics beyond the mentioned 10% error rates and the absence of comparison with other methods limit the depth of the argument. The discussion of industry practices (e.g., Elsevier’s pricing) is anecdotal but raises important considerations for researchers.

Scientific Rigor, Source Quality, Title Accuracy

The presentation is scientifically rigorous in its methodology, with a clear description of the annotation process and model development. The author does not cite specific external sources but refers to the use of PubMedBERT and mentions the ‘cat paper’ from Google, indicating familiarity with key AI developments. The title accurately reflects the content. The talk is based on the author’s own work, which adds credibility but also limits the generalizability. The mention of the NBRP workshop and the availability of a related video on YouTube provides additional context. Overall, the sources are appropriate for an expert opinion presentation, though more formal citations would enhance rigor.

The title accurately reflects the content, which details the development and implementation of an AI-based literature search system for DNA bank operations.

7/10

The presentation is based on the author's direct experience developing an AI system for literature screening in a DNA bank context. It provides specific details on methodology, data annotation, and performance metrics, but lacks external validation or peer-reviewed publication of the system. The claims about AI limitations and industry practices are anecdotal.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 00:00Introduction: Speaker introduces himself and the topic of AI-based literature search system for DNA bank.
• 02:00Explanation of AI development basics: history of AI winters and the 2012 breakthrough with deep learning (Google cat paper).
• 05:00Description of the supervised learning approach: annotation of 4000 papers from Scientific Reports by 7 team members.
• 08:00Details on the definition of genetic resources and the four categories used for annotation.
• 10:00Discussion of token limit (512 tokens) and the need to split papers into subsections (subsection method).
• 12:00Performance results: false positive and false negative rates around 10%, and speed of 100 papers per minute.
• 14:00Implementation in DNA bank operations: increase in overseas deposit requests from <10 to ~100 per year.
• 16:00Challenges with resource utilization papers: inconsistent author names and bank name variations, need for traceability.
• 18:00Issues with OpenAI filtering: suspicion that certain terms (e.g., 'sacrifice') are filtered, hiding relevant papers.
• 20:00Discussion of Elsevier's ScienceDirect AI service and the high cost of data access; Japan's lag in AI adoption.
• 22:00Conclusion: Emphasis on the necessity of using AI for literature exploration to avoid being left behind.

• NBRP Workshop 2025 — The presentation was part of this workshop.

• NBRP Workshop 2025 — The presentation was part of this workshop, providing context for the work.