引導式測驗 · 硬 等級
HyDE 假設文檔嵌入測驗
回顧 HyDE 生成、對比文件嵌入、零樣本檢索、樣本平均以及假設文本的限製作為證據。
問題 1 的 8
HyDE 在執行檢索之前會產生什麼?
預覽每個問題8 問題
當您參加測驗時,就會出現答案和解釋。
- HyDE 在執行檢索之前會產生什麼?
- SQL 查詢
- 僅同義詞列表
- 查詢的壓縮圖像
- 可以回答查詢的假設文檔
- What is the intended role of the encoder bottleneck when HyDE's hypothetical document contains invented details?
- Retain useful retrieval patterns while reducing some irrelevant specifics, without guaranteeing error removal
- Prove every generated sentence is factually correct
- Manually repair every fabricated claim before encoding
- Make fabricated details incapable of influencing any search result
- HyDE 的一個主要實際優勢是它的運作方式:
- 僅包含數百萬個帶標籤的範例
- 僅適用於英文文本
- 零樣本,無標記相關性訓練數據
- 僅用於圖片搜尋
- 原始 HyDE 中使用哪種類型的嵌入模型來對假設文檔進行編碼?
- 監督查詢分類器
- 無監督對比編碼器,例如 Contriever
- 基於規則的關鍵字擷取器
- 卷積圖像網絡
- HyDE 旨在解決密集檢索中的哪些核心問題?
- 簡短的查詢和較長的答案段落通常位於嵌入空間的不同區域
- 文檔太短,無法建立索引
- 倒排索引太慢
- 嵌入不能儲存在磁碟上
- How can a HyDE variant combine several generated hypothetical documents into one search representation?
- Declare all generated claims to be verified facts
- Delete the encoder and search only by file size
- Require every generated document to be a real corpus passage
- Average their encoded vectors before searching the real corpus
- Which text should support a factual answer after HyDE retrieval?
- The generated hypothetical document because generation verifies it
- Relevant real source passages that have been checked, rather than the hypothetical document alone
- Only the nearest vector's numeric coordinates
- Any invented detail repeated by several generated samples
- What extra retrieval cost can HyDE introduce compared with directly embedding a user's query?
- Mandatory relevance-label collection for every query
- A requirement to retrain Contriever for each search
- The latency and computation of generating one or more hypothetical documents
- Elimination of all encoder and index computation