Lesson 4 named the three pieces — blackboard, knowledge sources, control — and the property that forced them into existence: certainty has to flow both ways, not just bottom-up. This lesson opens the blackboard itself.
Hearsay-II’s board is organized into hypothesis levels, each a different grain of abstraction over the same stretch of audio:
- Segment — a slice of signal classified as roughly one phone (a vowel, a consonant sound).
- Syllable — segments grouped into a syllabic unit.
- Word — syllables grouped into a candidate word from the ~1,000-word lexicon.
- Phrase — words grouped into a candidate parse the grammar accepts.
Every hypothesis at every level carries a confidence and links to the hypotheses immediately above and below it that support it — a word hypothesis points down at the syllables that spell it and up at the phrases it could complete.
Bidirectional inference means a knowledge source can create or boost a hypothesis by reading either direction: a syllable KS proposes words from strong syllables (bottom-up), but a grammar KS can also propose which words are likely here from sentence context and go looking for their syllables in weak, ambiguous signal (top-down) — exactly the move a fixed pipeline structurally cannot make, because by the time the parser runs, the phoneme decision is already frozen.
That capability enables Hearsay-II’s signature search strategy, the island of certainty: don’t insist on resolving the utterance left to right. Find wherever, at whatever level, some KS is confident — the “island” — and schedule work that grows the interpretation outward from it in both directions, letting each newly secured hypothesis constrain its still-noisy neighbors. A strong island in the middle of a sentence is worth more than grinding through uncertain data at the edges first.
Given a confident word-level island appearing mid-utterance while the segment level at the start is still weak — what does the system actually do next?