Building and Understanding Ontology

Forte is built on top of an Ontology system, which defines the relations between NLP annotations, for example, the relation between words and documents, or between two words. This is the core for Forte.

The ontology can be specified via a JSON format. And tools are provided to convert the ontology into production code (Python). Make sure Forte is installed before following this tutorial.

A first glimpse of ontology generation

Imagine you need to develop an NLP system for a pet shop, first thing first, you need to understand what are the needed output from the documents. Let’s say you need to develop a system to assets such as Pet and Revenue , and hopefully automatically find out these from text. We have built an example ontology here: pet shop ontology

Now, before we go to the details, at Forte’s root directory, try run the following command:

generate_ontology create -i examples/ontology/pet_shop.json -o examples/ontology -r

If run successfully, you will find some python code being generated in examples/ontology, under the package This is what the Forte ontology system does, it generates the python classes needed to handle the NLP data structures.

The JSON ontology spec should be quite self-explanatory, we define types like Pet and Owner, which have some attributes. And the Owner have a list of Pet. The python code exactly represent the structure.

In the rest of the tutorial, we will walk through this example and you will learn:

  • Define a simple ontology spec for your project.

  • Import other ontology(s) to build yours.

  • Generate the corresponding Python classes automatically and use them in your project.

Before we start

There are a few basic concepts to understand Forte’s ontology system.

  • Entry - An entry corresponds to one NLP unit in the document, for instance, an annotated sequence or relationship between annotated sequences. Token, Sentence and DependencyLink are some examples of entries. One entry defined in the config is used to generate one python class.

  • Ontology - A collection of entries, where each entry is represented as a class. The entries of an ontology can span one or more Python modules. This way, the entry classes can be imported regardless of the ontology config they were originally generated from. The modules that contain these entry classes generally belong to the package ft.onto.

  • Attribute - An attribute generally corresponds to a label or property associated with an entry, like, pos_tag for the entry Token.

  • Top - Top entries are a set of entries that are pre-defined in the Forte library, in the module All user-defined entries should extend one of the top entries.

We provide a set of commonly used NLP entry types in the module Those entries could be used directly in your project!

A simple ontology config

Let us consider a simple ontology for documents of a pet shop.

    "name": "pet_shop_ontology",
    "description": "An Ontology Used to manage the pet shop assets and pets",
    "definitions": [
            "entry_name": "ft.onto.pet_shop.Pet",
            "parent_entry": "",
            "description": "Pets in the shop.",
            "attributes": [
                    "name": "pet_type",
                    "type": "str"
                    "name": "price",
                    "description": "Price for pet. A 2x2 matrix, whose columns are female/male and rows are juvenile/adult.",
                    "type": "NdArray",
                    "ndarray_dtype": "float",
                    "ndarray_shape": [2, 2]
            "entry_name": "ft.onto.pet_shop.Owner",
            "parent_entry": "",
            "description": "Owner of pets.",
            "attributes": [
                    "name": "name",
                    "type": "str"
                    "name": "pets",
                    "description": "List of pets the owner have.",
                    "type": "List",
                    "item_type": "ft.onto.pet_shop.Pet"

Breakdown of the simple ontology

  • The top level name and description are annotation keywords meant for descriptive purposes only.

  • The definitions is used to enlist entry definitions, where each entry is represented as a json object. Each entry correspond to one concept in the ontology, and a Python class.


Each definition is a dictionary of several keywords:

  • The entry_name keyword defines the name of the entry. It is used to define the full name of an entry, and is of the form <package_name><module_name>.<entry_name>.

    • The package name is generally ft.onto. It is used to create the package directory tree in which the generated module resides.

    • The <module_name> is the name of the generated file in which the entry would be placed.

    Note: Entries defined in the same config can have module names that are different from each other.

    • The <entry_name> is used as the generated class name.

  • The parent_type keyword defines the base class of the generated entry class. All the user-defined entries should inherit either any of the top entries or one of the other user-defined entries.

  • The description keyword is optionally used as the comment to describe the generated Python class.

  • attributes: List of attributes that would be used as instance variables of the generated class.


Each entry definition will define a couple (can be empty) attributes, mimicking the class variables:

  • The name keyword defines the name of the property unique to the entry.

  • The description keyword is optionally used as the comment to describe the attribute.

  • The type keyword is used to define the type of the attribute. Currently supported types are:

    • Primitive types - int, float, str, bool

    • Composite types - List, Dict, NdArray

    • Entries defined in the top module - The attributes can be of the type base entries (defined in the module) and can be directly referred by the class name.

    • User-defined types - The attributes can be of the type of entries that are user-defined. These user-defined entries could be defined (a) in the same config (b) any of the imported configs. To avoid ambiguity, only full-names of the user-defined entry types are supported

  • item_type: str: If the type of the property is a List, then item_type defines the type of the items contained in the list.

  • key_type and value_type: If the type of the property is a Dict, then these two represent the types of the key and value of the dictionary, currently, only primitive types are supported as the key_type.

  • ndarray_dtype: str and ndarray_shape: array: If the type of the property is a NdArray, then these two represent the data type and the shape of the array. NdArray allows storing a N-dimensional (N-d) array in an entry. For instance, through the simple ontology of pet shop above, we are able to instantiate Pet and name it dog. Then, we can assign a matrix to the attribute price by = [[2.99, 1.99], [4.99, 3.99]]. Internally, this $2 \times 2$ matrix is stored as a NumPy array. When ndarray_shape/ndarray_dtype is specified, the shape/data type of the upcoming array will be verified whether they match. If both ndarray_dtype and ndarray_shape are provided, a placeholder will be created by numpy.ndarray(ndarray_shape, dtype=ndarray_dtype).

Importing another ontology

imports is an optional keyword used to help you import existing ontology to help build the current one. This is similar to import in a normal programming language:

  • The entries of the imported configs can be used in the current config as types or parent classes.

  • The imports could either be

    • absolute paths

    • relative to the directory of the current config or the current working directory

    • relative to one of the user-provided spec_paths (see [generation steps](### ontology-generation-steps).)

For example, ft.onto.ft_module.Word has the parent entry defined in the generated module ft.onto.example_import_ontology. The generation framework makes sure that the imported JSON configs are generated before the current config. In case of cycle dependency between the JSON configs, an error would be thrown.

Package Naming Convention

Each entry should be named following a package convention, such as ft.onto.Pet in this example. This allows the generator to create a package structure for the python class.

In order to avoid polluting your package space accidentally, the package names that can be used on the types are restricted. The default package name allowed is ft.onto. However, in many cases you may want to use custom package name, let’s say,, how can we achieve it?

We can explicitly set more attributes in the additional_prefixes, as in the following snippet:

  "name": "pet_shop_ontology",
  "additional_prefixes": [
  "description": "An Ontology Used to manage the pet shop assets and pets",
  "definitions": [
      "entry_name": "",
      "parent_entry": "",
      "description": "Annotation for color words.",
      "attributes": [
          "name": "color_name",
          "type": "str"
      "entry_name": "",
      "parent_entry": "",
      "description": "Pets in the shop.",
      "attributes": [
          "name": "pet_type",
          "type": "str"
          "name": "color",
          "type": ""

Generating Python classes from ontology.

  • Write the json spec(s) as per the instructions in the previous sections.

  • Use the command generate_ontology --create (added during installation of Forte) to generate the ontology, and generate_ontology --clean to clean up the generated ontology. The steps are detailed in the following sections.

Ontology Generation Steps

At the beginning we have tried generating the ontology. Now let’s go into the some details.

  • To verify that the generate_ontology command is found, run generate_ontology -h, and the output should look like the following -

    $ generate_ontology -h
    usage: generate_ontology [-h] {create, clean} ...
    Utility to automatically generate or create ontologies.
    *create*: Generate ontology given a root JSON config.
    Example: python create --config forte/data/ontology/configs/example_ontology_config.json --no_dry_run
    *clean*: Clean a folder of generated ontologies.
    Example: python clean --dir generated-files
    positional arguments:
      {create, clean}
    optional arguments:
      -h, --help      show this help message and exit
  • All the arguments of generate_ontology create are explained as below:

    usage: generate_ontology create [-h] -i SPEC [-r] [-o DEST_PATH]
                                   [-s [SPEC_PATHS [SPEC_PATHS ...]]]
                                   [-m MERGED_PATH] [-e] [-a] [-l]
                                   [-n NAMESPACE_DEPTH]
    optional arguments:
     -h, --help            show this help message and exit
     -i SPEC, --spec SPEC  The main input JSON specification.
     -r, --no_dry_run      Generates the package tree in a temporary directory if true, ignores the argument `--dest_path`
     -o DEST_PATH, --dest_path DEST_PATH
                           Destination directory provided by the user. Only used when --no_dry_run is specified. The default directory is the current
                           working directory.
     -s [SPEC_PATHS ...], --spec_paths [SPEC_PATHS ...]
                           Paths in which the root and imported spec files are to be searched.
     -m MERGED_PATH, --merged_path MERGED_PATH
                           The destination file path for the mergedfile path.
     -a, --gen_all         If True, will generate all the ontology,including the existing ones shipped with Forte.
     -l, --lenient_prefix  If True, will not enforce prefix check.
     -n NAMESPACE_DEPTH, --namespace_depth NAMESPACE_DEPTH
                           set an integer argument namespace_depth to allow customized number of levels of namespace packaging.The generation of
                  for all the directory levels above namespace_depth will be disabled.Default value is set to 0 to disable
                           namespace packaging. When namespace_depth<=0, namespace packaging will be disabled and will be included in all
                           directory levels
  • Run the generate_ontology command in create mode. Let the destination path be a directory src in user project.

    $ generate_ontology create -i configs/example_complete_ontology_config.json -o src
    INFO:scripts.generate_ontology:Ontology will be generated in a temporary directory as -r is not specified by the user.
    INFO:scripts.generate_ontology:Ontology generated in the directory /var/folders/kw/ffm8tdn57vg3msn7fhr2htpw0000gp/T/tmpc1kp_xdp.
  • Note that the -o is not used at all and the data is generated in a temporary directory, as, by default, generate_ontology runs in the dry_run mode. To use the value of --dest_path, --no_dry_run has to be passed.

    $ generate_ontology create -i configs/example_complete_ontology_config.json -o src -r
    INFO:scripts.generate_ontology:Ontology generated in the directory /Users/mansi.gupta/user_project/src
  • Let’s look at the generated file structure.

    tree awesome
    └── pet

    Note that in this case we automatically generate a file in each level. With the -n option, we allow users to split the ontology definitions across multiple directories at customized level using namespace packaging. For example, if you wish to define another set of ontologies with prefix called at a different project folder but still want to access the entries defined in In this case you can generate the ontology file structure using generate_ontology create -n 1 which makes awesome a namespace package.

  • Our ontology generation is complete!

Cleaning the generated ontology

  • Use clean mode of generate_ontology to clean the generated files from a given directory.

  • All the arguments of generate_ontology clean are explained as below:

usage: generate_ontology clean [-h] --dir DIR [--force]

optional arguments:
 -h, --help  show this help message and exit
 --dir DIR   Generated files to be cleaned from the directory path.
 --force     If true, skips the interactive deleting offolders. Use with
  • Now, let’s try to clean up only the automatically generated files and directories. Say, there are user-created files in the generated folder, user_project/src/ft, called important_stuff that we do not want to clean up.

    $ mkdir user_project/src/ft/important_stuff
    $ touch user_project/src/ft/important_stuff/
    $ tree user_project
    ├── configs
    │   ├── example_complete_ontology_config.json
    │   └── example_import_ontology_config.json
    └── src
       ├── custom
          └── user
       └── ft
           ├── important_stuff
           └── onto
  • Run the cleanup command and observe the directory structure. The cleanup preserves the partial directory structure in the case there exists files that are not generated by the framework.

    $ generate_ontology clean --dir user-project/src
    INFO:scripts.generate_ontology.__main__:Directory /Users/mansi.gupta/user_project/src not empty, cannot delete completely.
    INFO:scripts.generate_ontology.__main__:Deleted files moved to /Users/mansi.gupta/user_project/.deleted/2019-11-28-02-53-29-952561.
  • For safety, the deleted directories are not immediately deleted but are moved to a timestamped directory inside .deleted and can be restored, unless --force is passed.

  • If the directories that are to be generated already exist, the files will be generated in the already existing directories.

  • Automatically generated folders are identified by an empty marker file of the name .generated, and automatically generated files are identified by special headers. If the headers or marker files are removed manually, than the cleanup won’t affect them.