uoe-nlp/multi3-nlu

Dataset Card for Multi³NLU++

Dataset Summary

Please access the dataset using

git clone https://huggingface.co/datasets/uoe-nlp/multi3-nlu/

Multi³NLU++ consists of 3080 utterances per language representing challenges in building multilingual multi-intent multi-domain task-oriented dialogue systems. The domains include banking and hotels. There are 62 unique intents.

Supported Tasks and Leaderboards

• multi-label intent detection
• slot filling
• cross-lingual language understanding for task-oriented dialogue

Languages

The dataset covers four language pairs in addition to the source dataset in English: Spanish, Turkish, Marathi, Amharic

Please find the source dataset in English here

Dataset Structure

Data Instances

Each data instance contains the following features: text, intents, uid, lang, and ocassionally slots and values

See the Multi³NLU++ corpus viewer to explore more examples.

An example from the Multi³NLU++ looks like the following:

{
        "text": "माझे उद्याचे रिझर्वेशन मला रद्द का करता येणार नाही?",
        "intents": [
            "why",
            "booking",
            "cancel_close_leave_freeze",
            "wrong_notworking_notshowing"
        ],
        "slots": {
            "date_from": {
                "text": "उद्याचे",
                "span": [
                    5,
                    12
                ],
                "value": {
                    "day": 16,
                    "month": 3,
                    "year": 2022
                }
            }
        },
        "uid": "hotel_1_1",
        "lang": "mr"

}

Data Fields

• 'text': a string containing the utterance for which the intent needs to be detected
• 'intents': the corresponding intent labels
• 'uid': unique identifier per language
• 'lang': the language of the dataset
• 'slots': annotation of the span that needs to be extracted for value extraction with its label and value

Data Splits

The experiments are done on different k-fold validation setups. The dataset has multiple types of data splits. Please see Section 4 of the paper.

Dataset Creation

Curation Rationale

Existing task-oriented dialogue datasets are 1) predominantly limited to detecting a single intent, 2) focused on a single domain, and 3) include a small set of slot types. Furthermore, the success of task-oriented dialogue is 4) often evaluated on a small set of higher-resource languages (i.e., typically English) which does not test how generalisable systems are to the diverse range of the world's languages. Our proposed dataset addresses all these limitations

Source Data

Initial Data Collection and Normalization

Please see Section 3 of the paper

Who are the source language producers?

The source language producers are authors of NLU++ dataset. The dataset was professionally translated into our chosen four languages. We used Blend Express and Proz.com to recruit these translators.

Personal and Sensitive Information

None. Names are fictional

Discussion of Biases

We have carefully vetted the examples to exclude the problematic examples.

Other Known Limitations

The dataset comprises utterances extracted from real dialogues between users and conversational agents as well as synthetic human-authored utterances constructed with the aim of introducing additional combinations of intents and slots. The utterances therefore lack the wider context that would be present in a complete dialogue. As such the dataset cannot be used to evaluate systems with respect to discourse-level phenomena present in dialogue.

Additional Information

Baseline models: Our MLP and QA models are based on the huggingface transformers library.

QA

We use the following code snippet for our QA experiments. Please refer to the paper for more details

https://github.com/huggingface/transformers/blob/main/examples/pytorch/question-answering/run_qa.py
python run_qa.py config_qa.json

Licensing Information

The dataset is Creative Commons Attribution 4.0 International (cc-by-4.0)

Citation Information

Coming soon

Contact

Nikita Moghe and Evgeniia Razumovskaia and Liane Guillou

Dataset card based on Allociné