Hyperlink Authority Propagation in Machine-Readable Knowledge Bases, Retrieval-Truncation Games in Generative Answering and High-SNR Snippet Extraction: chapter answers
When readers put their questions directly to a generative answering system, will the pages of a knowledge base be found, placed in the context and cited? Popular advice abounds, yet almost none of it says under what conditions it holds, how large its effect is, or what result would show it wrong. This book models, layer by layer, the path a piece of text travels before it is cited: authority on the link graph, lexical retrieval scores, truncation of the context, competition among content providers, and the reranker's choice. Its parameters are illustrative; all calculations are generated by i…
DOI pending · monograph page goes live after publication · 5DL-B-176 · 中文版
- Chapter 1 · Power Iteration and the Damping Factor: How Hyperlink Authority Flows Through a Knowledge Base
Chapter 1 computes PageRank and HITS with power iteration and a random surfer: in an illustrative knowledge base with damping factor 0.85, the mother page receives 0.3240, or 2.59 times the uniform share. - Chapter 2 · The Mother Page: How Internal-Link Topology Concentrates Authority
Chapter 2 proves that under star-shaped internal linking the mother page's share has a ceiling of $d/(1+d)$, or 0.4595, independent of the size of the base. - Chapter 3 · BM25 and Length Normalisation: the Term-Frequency Ledger of Stuffing, Padding and Direct Answers
Chapter 3 implements BM25 with the standard library and shows that stuffing soon peaks while padding is discounted by length: a twelve-word direct answer padded to two hundred and forty words keeps only 0.302 of its sco… - Chapter 4 · Top-k Truncation: the Probability that a Passage Enters the Context
Chapter 4 computes inclusion probabilities under top-k truncation: a ninety percent chance of the top three needs a score advantage of 2.46 standard deviations. - Chapter 5 · Competing for k Positions: Position Auctions, Homogenisation and the Commons
Chapter 5 borrows position auctions and ratio contests to show that ten competitors burn 0.900 of a position's value, and that systems which reward duplication homogenise content. - Chapter 6 · Signal-to-Noise and Softmax Reranking: the Selection Gain of Direct-Answer Snippets
Chapter 6, the core of the book, derives the selection gain of a direct-answer snippet in a softmax reranker: with twenty candidates at temperature one the gain is 4.65; structured elements lift the probability of selec… - Chapter 7 · The Classics Examined: Zheng Qiao's Classification and Bush's Associative Trails
Chapter 7 reads Zheng Qiao's Treatise on Collation beside Vannevar Bush's As We May Think, and shows with a program that classification helps retrieval only when its accuracy exceeds 0.667. - Chapter 8 · Practice and Falsification: the Copy Checker, a Publisher's Checklist and a Pre-Registered Test
Chapter 8 gives a publisher's checklist, a copy checker and a pre-registered test; detecting a moderate lift needs 1,081 runs per group. Marked boxes separate published evidence, runnable Python calculations (20 program…
Illustrative model calculation, not a live quote