mixsnips-intent-spans

MixSNIPS with per-intent segmentation, as character spans.

The source corpus marks slots with BIO tags and lists the intents of an utterance as an unordered set, but does not say which span of the utterance belongs to which intent. This dataset adds that missing structure: each utterance is segmented into contiguous per-intent spans, given as character offsets into text, with slot spans nested under the intent that owns them.

{
  "text": "book a diner for 1 in green isle and then will it be warmer at 15 o clock",
  "intents": [
    {"intent": "BookRestaurant", "start": 0, "end": 32,
     "slots": [{"name": "restaurant_type", "start": 7, "end": 12},
               {"name": "party_size_number", "start": 17, "end": 18},
               {"name": "city", "start": 22, "end": 32}]},
    {"intent": "GetWeather", "start": 42, "end": 73,
     "slots": [{"name": "condition_temperature", "start": 53, "end": 59},
               {"name": "timeRange", "start": 63, "end": 73}]}
  ],
  "source": "heuristic"
}

Offsets count Unicode codepoints into text — Python characters, not UTF-8 bytes or UTF-16 units — and are half-open, so text[start:end] is the span. Connective tokens ("and then") sit between segments, covered by no intent span.

Fields

Field	Description
text	The utterance, whitespace-joined source tokens
intents	Intent spans over text, each with the slot spans it contains
source	How the segmentation was derived: single, heuristic or llm

How The Segmentation Was Derived

The boundaries are reconstructed, not gold — the source corpus does not record them. source says how each row was settled, in decreasing order of confidence:

• single — a single-intent utterance, so the whole span is unambiguous.
• heuristic — slot types that pin down one intent fix the order, and the strongest connective in the gap between them places the seam.
• llm — the heuristics were not decisive, so a model chose the intent order and seam positions. Every answer was validated against the BIO spans before being accepted, but these rows are the least corroborated.

Treat the segmentation as approximate and spot-check before relying on it. Where the two methods overlap they agree closely, but there is no gold standard to measure either against.

Provenance

Utterances come from the SNIPS benchmark, released as CC0-1.0 in snipsco/nlu-benchmark, composed into multi-intent utterances by AGIF. Please cite the Snips and AGIF papers. The segmentation added here was produced with an open model served by Fireworks.

Produced by multi-nlu-data at revision c22b918.