מושגי ליבה
Large Language Models (LLMs) can effectively generate session data to enhance conversational search performance.
תקציר
ConvSDG proposes a framework using LLMs for session data generation in conversational search.
The framework explores dialogue-level and query-level data generation for fine-tuning conversational dense retriever.
Extensive experiments show ConvSDG outperforms baselines on widely used datasets.
Generated data improves system performance significantly, addressing data scarcity challenges.
Supervision signals from different query forms impact retrieval performance.
Varying sizes of generated data affect the effectiveness of fine-tuning in unsupervised and semi-supervised scenarios.
סטטיסטיקה
大規模言語モデル(LLM)を使用して、会話検索のパフォーマンスを向上させるためにセッションデータを生成する枠組みを提案します。