swap-uniba/EVWSD-ITA

This is the dataset card for EVWSD-ITA.

This data repo contains two files:

• —"ds_train.json": contains the train instances
• —"imgs.zip": contains the images associated to each train instance

Dataset Structure

Data Instances

{
    "id": "bn:00022412n",
    "hyp_id": "bn:00017670n",
    "gloss": "Atto del cuocere",
    "lemma": "cucina",
    "hyp_lemma": [
      "cambiamento di stato"
    ],
    "bns": [
      "bn:00018237n", ..., "bn:00049248n"
    ],
    "is_co_hyp": [
      true, ..., false
    ],
    "images": [
      "F14/bn:00018237n", ..., "F0/bn:00049248n"
    ],
    "all_lemmas": [
      ["masticazione", "masticare"],
        ...,
      ["cucine", "cucina", "cucina attrezzata", "cucina aperta", "cucinotto"]
    ],
    "all_glosses": [
      "La masticazione è il processo mediante il quale il cibo è frantumato e preparato dai denti.", ..., "Una stanza attrezzata per la preparazione dei cibi."
    ],
    "img": "F22/bn:00022412n"
}

Data Fields

All data fields are extracted from BabelNet

• —id: String, synset id for this instance
• —hyp_id: String, synset id for the hypernym of this istance
• —gloss: String, gloss in Italian for this instance
• —lemma: String, lemma in Italian for this instance
• —hyp_lemma: List of strings, all possible lemmas for the hypernym of this instance
• —bns: List of strings, all synset ids associated for this instance. The list will contain both: co-hyponyms of this instance, synsets that have the "lemma" of this instance as a possible lemma.
• —is_co_hyp: List of booleans, True if the corresponding synset in "bns" is a co-hyponym of this instance, False otherwise
• —images: List of strings, path to the image for each synset in "bns" in the directory extracted from the "imgs.zip" file
• —all_lemmas: List of lists of strings, each list contains all the Italian lemmas associated to each synset in "bns"
• —all_glosses: List of strings, all Italian glosses associated to each synset in "bns"
• —img: String, pato the the image associated to this instance

Important Notes

• —All images have been resized to 336x336, the same will be done for the test set.
• —One needs to create queries for each instance in this dataset. The test set will contain manually annotated queries by extracting words from: the instance lemma, the lemma of the hypernym and the instance gloss.
• —Using external data sources is allowed, since synsets that will be present in the test set may not be present in the train set, augmenting the dataset is recommended.
• —UPDATE: This train set and the future test set were created by leveraging BabelNet. To limit possible train-test set contamination, we ask the participants to not use BabelNet as data source for train set augmentation.