The system must decide whether a correction targets recent text or an arbitrary earlier span

What it is

“That word” and “the sentence just now” in a spoken correction are anaphors. The system has to decide whether the command aims at the most recent stretch (the hypothesis just emitted, the last sentence) or at an arbitrary earlier place in the document. A wrong scope edits a different sentence. This is reference resolution and edit scope, not the wording of the replace command itself.

Why it happens

People by default fix the error they just saw, so “change meeting to catch-up” prefers the latest hit. In long dictation the error may sit two paragraphs back, and they will say “the date in the third paragraph.” Without an explicit anchor the system can only use a recent window, occurrence counts of that word, and deictic phrases (“the line above,” “in the title”). Repeated words break the recency heuristic. Applying every edit at the end of the document is an implementation shortcut that leaves earlier errors unreachable. Failed scope feels like: the command was accepted, the wrong place did not move, a right place was damaged.

Studying it

Build documents with the target error in the last sentence, in the middle, or with the same word three times. Allow only spoken commands, no tapping. Measure whether the span that changed equals the intended span. Independent variables include whether phrases such as “previous sentence / sentence n” are allowed as anchors. Gaze or replay can check whether those phrases are actually used.

Where it stops holding

A one-sentence command has no history; scope is always the recent stretch. With a caret, scope can be defined as “from the caret,” and voice need not resolve the whole file. In a collaborative document “just now” may mean someone else’s edit, and temporal anaphora breaks. Treating a scope miss as a recognition miss sends work to ASR instead of to reference.

Applying it

• Default to the latest recognition or the current selection; to change earlier content, require a parseable anchor or that the selection be moved first.
• When the target word is not unique, list candidate occurrences before changing; do not silently edit the first or the last.
• Accept with “the error is in paragraph two and the command has no anchor” and “the same word three times,” and record wrong-place edits.

Related

• Same group: C7.13.1 Spoken correction depends on saying a replacement command rather than editing text directly · C7.13.3 Voice editing lacks the precise caret placement of keyboard editing · C7.13.4 Complex edits are less efficient to express in speech than by direct manipulation
• Adjacent: C7.10 Live Partial Results · C7.17 Switching Between Voice and Keyboard
• Search: correction scope · anaphora · edit target span