MASSIVE Multilingual NLU Dataset

## Overview MASSIVE (Multilingual Amazon SLURP-based corpus for Slot-filling, Intent classification, and Virtual-assistant Evaluation) is a parallel multilingual NLU dataset containing over 1 million utterances across 51 languages. Each utterance is annotated for intent prediction (60 intent classes) and slot annotation (55 slot types). The dataset was created by professional translators who localized the English SLURP dataset, which consists of single-shot interactions with an intelligent voice assistant. Format: text records with intent labels and span-level slot annotations. ## Schema - `id` — string — utterance identifier - `locale` — string — language-region code (e.g., en-US, ja-JP) - `partition` — string — train/dev/test split - `scenario` — string — high-level domain (e.g., alarm, music, weather) - `intent` — string — one of 60 intent classes (e.g., alarm_set) - `utt` — string — the localized utterance - `annot_utt` — string — utterance with inline slot annotations - `worker_id` — string — anonymized annotator id - `slot_method` — list — per-slot localization method (translation, transcreation, etc.) - `judgments` — list — quality judgments from reviewers ## Sources - AmazonScience/massive on HuggingFace — https://huggingface.co/datasets/AmazonScience/massive — CC-BY-4.0 - Original paper: arXiv:2204.08582 - Derived from SLURP (Bastianelli et al., 2020) ## Methodology Amazon Science started from the English SLURP voice-assistant dataset and engaged professional native-speaker translators to localize each utterance into 50 additional languages, preserving intent labels and re-annotating slots in the target language. Translators used either direct translation or transcreation (cultural adaptation) per slot, and the dataset records which method was used. Multiple judgments per item were collected for quality control. Train/dev/test partitions are parallel across all 51 languages. ## Known gaps & limitations Utterances are short, single-turn voice-assistant queries — coverage of long-form or multi-turn dialogue is out of scope. The intent/slot schema reflects a 2020-era smart-speaker product surface (alarms, music, IoT, weather) and may not generalize to newer assistant capabilities. Some low-resource languages have higher translation-noise rates flagged in the source paper. Locale coverage is uneven for dialectal variation (e.g., only zh-CN and zh-TW for Chinese). Buyers should validate label consistency empirically for any specific locale before fine-tuning. ## Intended use & out-of-scope - **IS for**: training and evaluating multilingual intent classification and slot-filling models, cross-lingual transfer benchmarking, multilingual NLU fine-tuning for voice/chat assistants, and as an eval set for instruction-following agents across languages. - **NOT for**: open-domain or multi-turn dialogue modeling, ASR/speech tasks (text-only), or benchmark training of models that will later be evaluated on MASSIVE itself (leakage risk — widely used eval). _Federated dataset: 162 parquet shards, 189.8 MB total. Queries and downloads stream through the DataBazaar API._ Original supplier listing: MASSIVE: Multilingual NLU Dataset (51 Languages, 1M+ Utterances) Parallel multilingual NLU benchmark from Amazon Science with 1M+ utterances across 51 languages, annotated with 60 intents and 55 slot types. Built by localizing SLURP voice assistant interactions.