Finding the Most Compute-Optimal Recipe for Repurposing Language Models into Embedding Models
This research paper presents an algorithm for determining the optimal combination of model size, data quantity, and fine-tuning method to create high-quality text embedding models from pre-trained language models while adhering to specific computational budgets.
基於詞組對齊偏好的機器翻譯
本文提出了一種基於詞組對齊偏好的方法，用於減輕基於大型語言模型的機器翻譯系統中的幻覺和遺漏現象。
단어 정렬 선호도를 활용한 기계 번역에서의 오역 및 누락 문제 완화
단어 정렬 정보를 활용한 선호도 최적화를 통해 LLM 기반 기계 번역 모델의 고질적인 문제인 오역 및 누락 현상을 완화할 수 있다.
Mitigating Hallucination and Omission in LLM-Based Machine Translation Using Word Alignment Preference
Leveraging word alignment as a preference signal during optimization can effectively reduce hallucination and omission errors in Large Language Model (LLM)-based machine translation systems.
線上社群中的人類價值觀研究
本文提出了一種利用自然語言處理技術，自動化分析 Reddit 等線上社群中所展現的人類價值觀的方法，並探討其與傳統問卷調查結果的關聯性。
온라인 커뮤니티에서 나타나는 인간 가치관 분석: 레딧을 활용한 대규모 연구
본 논문에서는 레딧에서 나타나는 인간 가치관을 대규모로 분석하여 온라인 커뮤니티 연구에 활용할 수 있는 방법론을 제시합니다.
Enhancing Open-Domain Question Answering with REAR: A Relevance-Aware Retrieval-Augmented Framework
REAR, a novel framework for open-domain question answering, improves the accuracy and reliability of retrieval-augmented generation (RAG) systems by incorporating a relevance-aware architecture and specialized training methods to enhance the model's ability to assess and utilize retrieved documents effectively.
オンラインコミュニティにおける人間の価値観の分析：大規模テキストデータを用いた価値観抽出モデルの開発と評価
本稿では、オンラインコミュニティにおける人間の価値観を大規模に分析するため、Redditの投稿データからシュワルツの価値観を抽出するモデルを開発し、その有効性を検証しました。
Extracting and Analyzing Schwartz's Human Values from Reddit Communities
This research leverages NLP to extract and analyze human values within Reddit communities, revealing insights into online behavior and demonstrating the potential of computational methods for social science research.
PaDeLLM-NER：大型語言模型中用於命名實體識別的平行解碼方法
本文提出了一種名為 PaDeLLM-NER 的新型平行解碼方法，用於加速大型語言模型在命名實體識別任務中的推理速度，並在保持高預測準確率的同時，顯著縮短了序列長度和推理時間。
