Subject-dropped fragments need completing from the last turn

What it is

The system asks “black tea or green.” The user says “green.” The next line is two syllables: “hot.” No verb, no object, no subject. This is not a skill named “hot.” It grafts temperature onto the order from the last turn. Ellipsis completion attaches a fragmentary utterance to the previous predicate and restores a full proposition. Chinese especially runs on zero anaphora: the dropped piece is grown back from the last frame. The object of work is a syntactically incomplete follow-up, not finding a noun for “it.”

Why it happens

Conversational ellipsis leaves slots recoverable from the prior predicate. Run the surface string through intent alone and it becomes another skill: “hot” sets the HVAC, “two” sets a timer, “no sugar” falls out of domain. Completion copies the previous user or system frame, overlays the fragment on the matching slot (temperature = hot), then executes that frame. Fragments are the majority of follow-ups in slot filling, not a corner case.

The difference from pronoun resolution is that often there is no pronoun. There is no “it” to bind, only a modifier or an alternative label. A system that admits a follow-up only when the sentence is complete will treat the most common spoken shape as a new session. Completion has to run before, or beside, intent classification: first “does this look like a fragment of the last frame,” then “is this a new intent.”

Studying it

Collect follow-up fragments. Hand-split them into: elliptical answers to alternative questions, modifiers of the prior order, corrections, and true new intents. Compare intent error and slot error when NLU sees the fragment alone versus the fragment concatenated with the previous turn. Languages that drop subjects (Chinese, Japanese) should report fragment rate on their own; do not estimate it from English full-clause corpora.

Turn-distance between fragment and prior turn can be an independent variable — only to see whether completion still works, not as a study of how long context lives. Read scripts almost never produce real ellipsis; elicit talk.

Where it stops holding

Once the user has wrenched the topic elsewhere (“green,” then immediately “play a song”), the fragment rule will glue a new intent back onto tea. A full repeat of the choice (“I’ll take hot green tea”) does not need completion; forcing it stacks structure. Words that look like skill names (“pause,” “off”), even in follow-up position, may be new device control and must not always overlay the last frame. A first utterance with no prior turn cannot be ellipsis.

Applying it

• When the parse has no verb, or the whole utterance is a slot value, try an overlay on the previous frame before a new-intent path. Ask only if both score high.
• Speak the completed proposition in the reply (“hot green tea”) so a wrong bind can be heard and repaired.
• Count “fragment taken as new intent.” That is this rule’s main failure, not word error.
• Regression on real order or shopping follow-ups: last turn “green,” this turn “hot / large / less sugar” must remain one order.

Related

• Same group: M1.09.1 Referent candidates come from talk, screen, and the room · M1.09.3 New-turn slots need overwrite versus overlay
• Nearby: M1.04 Context Retention · M2.13 Multi-Intent and Compound Commands · M2.07 Dialogue flow and state design
• Search terms: ellipsis completion · zero anaphora · fragment follow-up