← All posts

· Casimiro Ferreira· 2 min read

A Multilingual Sentence-Types Dataset: Questions, Commands, Statements

  • Datasets
  • Multilingual
  • NLP
  • Intent
  • Classification
  • FOSS

A voice assistant’s routing logic depends on knowing what kind of sentence it received before it tries to answer anything. A question needs an answer. A command needs execution. A statement might need acknowledgment or storage. Getting that classification right, in any language the user speaks, is the prerequisite for everything else.

sentence-types-multilingual is the training corpus behind that layer: 69,300 labeled sentences, 9,900 for each of seven languages (English, Spanish, French, German, Italian, Portuguese, and Dutch).

What the labels mean in practice

The dataset uses a flat set of six labels: one label column per row, with no type/sub-type split. These labels map directly to how little_questions (the inference library that consumes this data) routes utterances:

  • wh_question: questions built around a wh-word (what, where, who, and so on).
  • polar_question: yes/no questions. The EAT (Expected Answer Type) taxonomy inside little_questions adds 53 fine-grained answer-type labels (person, location, quantity, definition, and so on) on top of the question labels, but sentence-type classification is the first gate.
  • command: imperative forms. Commands don’t expect an answer. They expect an action.
  • request: polite or indirect asks for action, distinct from a bare imperative.
  • statement: declarative. Statements in a dialogue context often carry polarity that matters downstream. A yes/no/maybe classifier runs on statements to interpret answers to prior questions.
  • exclamation: emotionally marked utterances that need different handling than neutral declaratives.
{
  "language": "en",
  "label": "wh_question",
  "text": "What time is it?"
}

Why cross-linguistic coverage is non-trivial

The same communicative intent surfaces differently in different grammars:

  • English marks questions with word-order inversion. Portuguese and Spanish often mark them with punctuation and intonation alone, leaving word order untouched.
  • German separates verbs to sentence-final position in ways that shift where the classifying signal lives.
  • Romance languages use dedicated imperative morphology for commands that English expresses with the bare verb.

A model trained only on English gets these wrong everywhere else. Parallel labeled data across the seven languages provides the cross-lingual signal per-language classifiers need. The same generation pipeline extends to further languages as they are added.

The downstream stack

The models trained on this data ship inside little_questions, a zero-dependency offline library (numpy + onnxruntime) with per-language ONNX classifiers for sentence type and a 43-language yes/no polarity model. Models are bundled in-wheel for English and lazy-downloaded for other languages. The sentence-type classifiers are published as TigreGotico/sentence-types on HuggingFace. The EAT answer-type classifiers are trained internally and not publicly released.

from little_questions import Sentence

s = Sentence("What time is it?")
print(s.sentence_type)     # "question"
print(s.classification)    # e.g. "NUM:date"

little_questions is the natural-language routing layer for OVOS and LILACS: classifying whether an utterance is a question, command, or statement is the first dispatch decision a voice pipeline makes.

sentence-types-multilingual on HuggingFace · little_questions on GitHub