What are metadata ?

Metadata are any information about the dataset that are not directly part of the data.

General metadata about an entire dataset specify things like:

  • Who are the authors
  • What is the name of the dataset
  • How should it be cited
  • What other datasets it is related to or derived from
  • What is its identifier (DOI)
  • Under what license it is shared
  • Relevant keywords

This information is usually provided in introductions or documentation files, in a way that is easy to understand for humans. To make it accessible automatically, it needs to be specified more formally. This is what the <dataset>.package.json file in this standard is for (see frictionless metadata).

Many other pieces of information are often left implicit, as it is easy for humans, given enough context, to guess them. However, this is neither future-proof (we will lose part of the context) nor machine-readable (software is terrible at picking up informally expressed context). These piece of information are:

  • What are the tables present in the dataset ?
  • What are the relations between tables ?
  • What are the columns present in the tables ?
  • What do these columns mean ?
    • This can be documented through linking to specific ontologies and vocabularies. Eg. a "glottocode" column may link to language codes in glottolog, a 'UD' column may provide cell or feature definitions following the Universal Dependencies conventions.
  • What should we expect to find in these columns ?
    • For example, what is the type of information present ? one column might only have numbers, another text, yet another only has binary true/false values, etc.
    • How are missing data expressed ? What does an empty cell mean ?
    • Are there constraints on these values ? For example, a minimum value for numbers, a set of characters used in phonological transcription, values which must be unique to each row, or mandatory.
  • Which column serves as identifiers ?
    • Identifiers are unique values which serve to refer unambiguously to a specific row of a specific table.

To record this metadata in a standardised, formal way, we use the frictionless standard to write it in a json file. For example, a json file expressing only information about data contributors might look like this:

  "contributors": [{
    "title": "Joe Bloggs",
    "email": "joe@bloggs.com",
    "path": "http://www.bloggs.com",
    "role": "author"

Because there is much more information to specify, metadata tends to be much longer, and is not practical to write by hand. We provide a python package paralex which can auto-generate metadata for you, because it has information (context !) about standard tables and columns names. For more on how to use it, see the tutorial.