MASSIVE

News

• 20 Apr 2022: Launch and release of the MASSIVE dataset, this repo, the MASSIVE paper, the leaderboard, and the Massively Multilingual NLU 2022 workshop and competition.

Quick Links

• MASSIVE paper
• MASSIVE Leaderboard and Massively Multilingual NLU 2022 Competition
• MASSIVE Blog Post

Introduction

MASSIVE is a parallel dataset of > 1M utterances across 51 languages with annotations for the Natural Language Understanding tasks of intent prediction and slot annotation. Utterances span 60 intents and include 55 slot types. MASSIVE was created by localizing the SLURP dataset, composed of general Intelligent Voice Assistant single-shot interactions.

Accessing and Processing the Data

The dataset can be downloaded here.

$ curl https://amazon-massive-nlu-dataset.s3.amazonaws.com/amazon-massive-dataset-1.0.tar.gz --output amazon-massive-dataset-1.0.tar.gz
$ tar -xzvf amazon-massive-dataset-1.0.tar.gz
$ tree 1.0
1.0
├── LICENSE
└── data
    ├── af-ZA.jsonl
    ├── am-ET.jsonl
    ├── ar-SA.jsonl
    ...

The dataset is organized into files of JSON lines. Each locale (according to ISO-639-1 and ISO-3166 conventions) has its own file containing all dataset partitions. An example JSON line for de-DE has the following:

{
  "id": "0",
  "locale": "de-DE",
  "partition": "test",
  "scenario": "alarm",
  "intent": "alarm_set",
  "utt": "weck mich diese woche um fünf uhr morgens auf",
  "annot_utt": "weck mich [date : diese woche] um [time : fünf uhr morgens] auf",
  "worker_id": "8",
  "slot_method": [
    {
      "slot": "time",
      "method": "translation"
    },
    {
      "slot": "date",
      "method": "translation"
    }
  ],
  "judgments": [
    {
      "worker_id": "32",
      "intent_score": 1,
      "slots_score": 0,
      "grammar_score": 4,
      "spelling_score": 2,
      "language_identification": "target"
    },
    {
      "worker_id": "8",
      "intent_score": 1,
      "slots_score": 1,
      "grammar_score": 4,
      "spelling_score": 2,
      "language_identification": "target"
    },
    {
      "worker_id": "28",
      "intent_score": 1,
      "slots_score": 1,
      "grammar_score": 4,
      "spelling_score": 2,
      "language_identification": "target"
    }
  ]
}

id: maps to the original ID in the SLURP collection. Mapping back to the SLURP en-US utterance, this utterance served as the basis for this localization.

locale: is the language and country code accoring to ISO-639-1 and ISO-3166.

partition: is either train, dev, or test, according to the original split in SLURP.

scenario: is the general domain, aka "scenario" in SLURP terminology, of an utterance

intent: is the specific intent of an utterance within a domain formatted as {scenario}_{intent}

utt: the raw utterance text without annotations

annot_utt: the text from utt with slot annotations formatted as [{label} : {entity}]

worker_id: The obfuscated worker ID from MTurk of the worker completing the localization of the utterance. Worker IDs are specific to a locale and do not map across locales.

slot_method: for each slot in the utterance, whether that slot was a translation (i.e., same expression just in the target language), localization (i.e., not the same expression but a different expression was chosen more suitable to the phrase in that locale), or unchanged (i.e., the original en-US slot value was copied over without modification).

judgments: Each judgment collected for the localized utterance has 6 keys. worker_id is the obfuscated worker ID from MTurk of the worker completing the judgment. Worker IDs are specific to a locale and do not map across locales, but are consistent across the localization tasks and the judgment tasks, e.g., judgment worker ID 32 in the example above may appear as the localization worker ID for the localization of a different de-DE utterance, in which case it would be the same worker.

intent_score : "Does the sentence match the intent?"
  0: No
  1: Yes
  2: It is a reasonable interpretation of the goal

slots_score : "Do all these terms match the categories in square brackets?"
  0: No
  1: Yes
  2: There are no words in square brackets (utterance without a slot)

grammar_score : "Read the sentence out loud. Ignore any spelling, punctuation, or capitalization errors. Does it sound natural?"
  0: Completely unnatural (nonsensical, cannot be understood at all)
  1: Severe errors (the meaning cannot be understood and doesn't sound natural in your language)
  2: Some errors (the meaning can be understood but it doesn't sound natural in your language)
  3: Good enough (easily understood and sounds almost natural in your language)
  4: Perfect (sounds natural in your language)

spelling_score : "Are all words spelled correctly? Ignore any spelling variances that may be due to differences in dialect. Missing spaces should be marked as a spelling error."
  0: There are more than 2 spelling errors
  1: There are 1-2 spelling errors
  2: All words are spelled correctly

language_identification : "The following sentence contains words in the following languages (check all that apply)"
  1: target
  2: english
  3: other
  4: target & english
  5: target & other
  6: english & other
  7: target & english & other

Note that the en-US JSON lines will not have the slot_method or judgment keys, as there was no localization performed. The worker_id key in the en-US file corresponds to the worker ID from SLURP.

{
  "id": "0",
  "locale": "en-US",
  "partition": "test",
  "scenario": "alarm",
  "intent": "alarm_set",
  "utt": "wake me up at five am this week",
  "annot_utt": "wake me up at [time : five am] [date : this week]",
  "worker_id": "1"
}

Preparing the Data in datasets format (Apache Arrow)

The data can be prepared in the datasets Apache Arrow format using our script:

python scripts/create_hf_dataset.py -d /path/to/jsonl/files -o /output/path/and/prefix

If you already have number-to-intent and number-to-slot mappings, those can be used when creating the datasets-style dataset:

python scripts/create_hf_dataset.py \
    -d /path/to/jsonl/files \
    -o /output/path/and/prefix \
    --intent-map /path/to/intentmap \
    --slot-map /path/to/slotmap

Training an Encoder Model

We have included intent classification and slot-filling models based on the pretrained XLM-R Base or mT5 encoders coupled with JointBERT-style classification heads. Training can be conducted using the Trainer from transformers.

We have provided some helper functions in massive.utils.training_utils, described below:

• create_compute_metrics creates the compute_metrics function, which is used to calculate evaluation metrics.
• init_model is used to initialize one of our provided models.
• init_tokeinzer initializes one of the pretrained tokenizers.
• prepare_collator prepares a collator with user-specified max length and padding strategy.
• prepare_train_dev_datasets, which loads the datasets prepared as described above.
• output_predictions, which outputs the final predictions when running test.

Training is configured in a yaml file. Examples are given in examples/. A given yaml file fully describes its respective experiment.

Once an experiment configuration file is created, training can be performed using our provided training script. We also have provided a conda environment configuration file with the necessary dependencies that you may choose to use.

conda env create -f conda_env.yml
conda activate massive

Then run training:

scripts/train.py -c YOUR/CONFIG/FILE.yml

Distributed training can be run using torch.distributed.launch. For example:

python -m torch.distributed.launch --nproc_per_node=8 scripts/train.py -c YOUR/CONFIG/FILE.yml

Seq2Seq Model Training

Sequence-to-sequence (Seq2Seq) model training is performed using the MASSIVESeq2SeqTrainer class. This class inherits from Seq2SeqTrainer from transformers. The primary difference with this class is that autoregressive generation is performed during validation, which is turned on using the predict_with_generate training argument. Seq2Seq models use teacher forcing during training.

For text-to-text modeling, we have included the following functions in massive.utils.training_utils:

• convert_input_to_t2t
• convert_intents_slots_to_t2t
• convert_t2t_batch_to_intents_slots

For example, mT5 Base can be trained on an 8-GPU instance as follows:

python -m torch.distributed.launch --nproc_per_node=8 scripts/train.py -c experiments/mt5_base_t2t_20220411.yml 2>&1 | tee /PATH/TO/LOG/FILE

Performing Inference on the Test Set

Test inference requires a test block in the configuration. See examples/xlmr_base_test_20220411.yml for an example. Test inference, including evaluation and output of all predictions, can be executed using the scripts/test.py script. For example:

python -m torch.distributed.launch --nproc_per_node=8 scripts/test.py -c experiments/xlmr_base_test_20220411.yml 2>&1 | tee /PATH/TO/LOG/FILE

Be sure to include a test.predictions_file in the config to output the predictions.

For official test results, please upload your predictions to the eval.ai leaderboard.

Hyperparameter Tuning

Hyperparameter tuning can be performed using the Trainer from transformers. Similarly to training, we combine all configurations into a single yaml file. An example is given here: example/xlmr_base_hptuning_20220411.yml.

Once a configuration file has been made, the hyperparameter tuning run can be initiated using our provided scripts/run_hpo.py script. Relative to train.py, this script uses an additional function called prepare_hp_search_args, which converts the hyperparameter search space provided in the configuration into an instantiated ray search space.

Licenses

See LICENSE.txt, NOTICE.md, and THIRD-PARTY.md.

Citation

We ask that you cite both our MASSIVE paper and the paper for SLURP, given that MASSIVE used English data from SLURP as seed data.

MASSIVE paper:

@misc{fitzgerald2022massive,
      title={MASSIVE: A 1M-Example Multilingual Natural Language Understanding Dataset with 51 Typologically-Diverse Languages}, 
      author={Jack FitzGerald and Christopher Hench and Charith Peris and Scott Mackie and Kay Rottmann and Ana Sanchez and Aaron Nash and Liam Urbach and Vishesh Kakarala and Richa Singh and Swetha Ranganath and Laurie Crist and Misha Britan and Wouter Leeuwis and Gokhan Tur and Prem Natarajan},
      year={2022},
      eprint={2204.08582},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

SLURP paper:

@inproceedings{bastianelli-etal-2020-slurp,
    title = "{SLURP}: A Spoken Language Understanding Resource Package",
    author = "Bastianelli, Emanuele  and
      Vanzo, Andrea  and
      Swietojanski, Pawel  and
      Rieser, Verena",
    booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
    month = nov,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2020.emnlp-main.588",
    doi = "10.18653/v1/2020.emnlp-main.588",
    pages = "7252--7262",
    abstract = "Spoken Language Understanding infers semantic meaning directly from audio data, and thus promises to reduce error propagation and misunderstandings in end-user applications. However, publicly available SLU resources are limited. In this paper, we release SLURP, a new SLU package containing the following: (1) A new challenging dataset in English spanning 18 domains, which is substantially bigger and linguistically more diverse than existing datasets; (2) Competitive baselines based on state-of-the-art NLU and ASR systems; (3) A new transparent metric for entity labelling which enables a detailed error analysis for identifying potential areas of improvement. SLURP is available at https://github.com/pswietojanski/slurp."
}