QubitPi/Antiqua

Antiqua Datasets

from datasets import load_dataset
dataset = load_dataset("QubitPi/Antiqua")

If dataset = load_dataset("QubitPi/Antiqua") throws an error, please upgrade the datasets package to its latest version

There are 4 splits we can choose from:

1. German

   split = dataset["German"]
   

2. German

   split = dataset["Italian"]
   

3. Latin

   split = dataset["Latin"]
   

4. Ancient Greek

   split = dataset["AncientGreek"]
   

To iterate the split for each triple of source node, target node, and link:

graph = split.iter(batch_size=1)
for triple in graph:
    source_node = {k: v for k, v in triple["source"][0].items() if v}
    target_node = {k: v for k, v in triple["target"][0].items() if v}
    link        = triple["link"][0]

Each source_node have the following fields (those in bold exist in all nodes while others are not always present)

• label: the vocabulary term

• language: the language of this term and is always one of the following values:

  • German
  • Latin
  • Ancient Greek

• audio: the link to the Wiktionary audio of this vocabulary

• declension-i-j: The declension entry in the i-th row and j-th column in declension table

  To encode the inflections which are common in most Indo-European languages, a virtual inflection table that looks like the following are employed (taking German noun "Gegenstand" as an example):

  ["",         singular,                    plural      ]
  [nominative, Gegenstand,                  Gegenstände ]
  [genitive,   "Gegenstandes, Gegenstands", Gegenstände ]
  [dative,     Gegenstand,                  Gegenständen]
  [accusative, Gegenstand,                  Gegenstände ]
  

  The declension (inflection) table above is equivalent to

  	singular	plural
  nominative	Gegenstand	Gegenstände
  genitive	Gegenstandes, Gegenstands	Gegenstände
  dative	Gegenstand	Gegenständen
  accusative	Gegenstand	Gegenstände

  This has been encoded in Antiqua Hugging Face Datasets as:

  • declension-0-0 = ""
  • declension-0-1 = "singular"
  • declension-0-2 = "plural"
  • declension-1-0 = "nominative"
  • declension-1-1 = "Gegenstand"
  • declension-1-2 = "Gegenstände"
  • declension-2-0 = "genitive"
  • declension-2-1 = "Gegenstandes, Gegenstands"
  • ...
  • declension-4-2 = "Gegenstände"

For example, to get the word and language of this word:

graph = split.iter(batch_size=1)
for triple in graph:
    source_node = {k: v for k, v in triple["source"][0].items() if v}
    print("label: {label}".format(label=source_node["label"]))
    print("language: {language}".format(language=source_node["language"]))

Each target_node is either

• an object with the same field scheme as source_node, or
• a "definition" node with a non-null "label" field only, in which case, the "label" field is the English meaning of the word represented by the source_node pointing at this target_node

graph = split.iter(batch_size=1)
for triple in graph:
    target_node = {k: v for k, v in triple["target"][0].items() if v}
    if target_node["language"] is None:
        print("This is an English definition of some source node word: {definition}".format(definition=target_node["label"]))
    else:
        print("label: {label}".format(label=target_node["label"]))
        print("language: {language}".format(language=target_node["language"]))

In the case of a "definition" node, we will see all fields are null except for the "label" field. Why are there so many unnecessary null-fields? Because there are decent amount fo target nodes that are not "definition" nodes; these are source nodes themselves acting as target nodes because they have been spotted to have connections with other words

Each link, no surprises there, encodes the relationship between the source_node and target_node. Its label field signifies the nature of the relationship. For example, { "label": "definition" } means that target_node is the definition of source_node

graph = dataset[split].iter(batch_size=1)
for triple in graph:
    link = triple["link"][0]
    print(link["label"])