Visualization tool for NLP information extraction: Named entity recognition, Entity attribute extraction, and Relation extration.
The ie-viz is a lightweight tool for in-line information extraction (IE) visualization. It supports customizable colors for named entity marking, tooltip for entity attributes display, and relation path that links related named entities. Current version has built-in light and dark themes while customization is available through the CSS. The ie-viz can be deployed as a Flask App that runs on a host:port. It can also be rendered as HTML and display in Browser or Interactive Python environments (e.g., Jupyter Notebook).
| Features | Support |
|---|---|
| Named Entity Marks | ✅ with customizable colors and table view |
| Entity Attributes | ✅ as tooltip and table view |
| Entity Relations | ✅ as path linking entities and table view |
| Filtering | ✅ by entity types |
| Theme | ✅ light and dark themes |
| Deployment | ✅ Flask APP or HTML rendering |
The filtering feature supports OR and AND logic for all available entity attributes. The table panel (collapsible) displays the selected entities, attributes, and relations.
- python ^3.11
- flask >=2.3
Python package is available on PyPi.
pip install ie-vizWe use a short sample News article synthesized by GPT-4o to demo this quick-start.
text = """
On Monday, the tech giant, Innovia Corp, announced the launch of its new flagship smartphone, the Innovia XPro.
The device, which comes in both 128GB and 256GB storage variants, is powered by the latest Octa-core Quantum processor
and features a 6.5-inch AMOLED display. According to the CEO, Lisa Martin, this release marks a significant milestone for the company.
The phone is expected to compete directly with the recent release from its rival, Nexon Technologies, which unveiled the
Nexon Ultra series earlier this year.
"""Entities have attribute type as one of "organization", "person", or "product".
- when type is organization, there is no other attributes
- when type is person, the entity has an optional Role attribute
- when type is product, the entity has an optional Specifications attribute
entities = [
{'entity_id': '0', 'start': 28, 'end': 40, 'attr': {'Type': 'organization'}},
{'entity_id': '1', 'start': 99, 'end': 111, 'attr': {'Type': 'product', 'Specifications': '128GB, 256GB, Octa-core Quantum processor, 6.5-inch AMOLED display'}},
{'entity_id': '2', 'start': 296, 'end': 307, 'attr': {'Type': 'person', 'Role': 'CEO'}},
{'entity_id': '3', 'start': 452, 'end': 470, 'attr': {'Type': 'organization'}},
{'entity_id': '4', 'start': 492, 'end': 510, 'attr': {'Type': 'product'}}
]Relations (optional) are listed as as entity id pairs.
relations = [
{'entity_1_id': '0', 'entity_2_id': '1'},
{'entity_1_id': '0', 'entity_2_id': '2'},
{'entity_1_id': '3', 'entity_2_id': '4'}
]The serve() function starts a Flask App that visualizes the entities, attributes, and relations. The theme parameter supports light and dark modes. The color_attr_key parameter specifies that entity color is based on entity "Type" attribute.
from ie_viz import serve
serve(text, entities, relations, theme="light", color_attr_key="Type")A Flask App starts on the localhost port 5000 (default). The full rendered HTML is available TechNews
* Serving Flask app 'ie_viz.utilities'
* Debug mode: off
WARNING: This is a development server. Do not use it in a production deployment. Use a production WSGI server instead.
* Running on all addresses (0.0.0.0)
* Running on http://127.0.0.1:5000
Press CTRL+C to quit
127.0.0.1 - - [02/Oct/2024 11:44:58] "GET / HTTP/1.1" 200 -
Below are Jupyter notebooks that demo some use cases:
Tech News: Organization, Product, and Person
Clinical Note: Medication, condition, and Adverse Reaction
Input formats
Both the serve() and render() functions accept the same data types for input.
The entities must be a list of dictionaries. Each dictionary must have entity_id, start, and end keys. The attr key is optional and can be used for entity attributes display or entity coloring.
entities = [
{'entity_id': '<entity id>', 'start': <start char>, 'end': <end char>},
{'entity_id': '<entity id>', 'start': <start char>, 'end': <end char>, 'attr': {'<attribute key>': '<attribute value>'}},
...
]The relations is optional. It must be a list of dictionaries with entity_1_id and entity_2_id keys.
relations = [
{'entity_1_id': '<entity id>', 'entity_2_id': '<entity id>'},
{'entity_1_id': '<entity id>', 'entity_2_id': '<entity id>'},
...
]Entity colors
There are two parameters to customize the entity colors, color_attr_key or color_map_func. If none of them are defined, a default color is assigned to all entities.
The color_attr_key is a easier way to define entity color. It specifies an attribute key to be used. All entities with the same attribute value will be assigned the same color.
from ie_viz import serve
serve(text, entities, color_attr_key="<attribute key to assign color>")Note that all entities must have that attribute key, or an error will be raised.
The color_map_func is a more flexible way to define colors. Users define a custom function that inputs an entity and returns a default color name or a hex color code (as string).
def color_map_func(entity) -> str:
if entity['attr']['<attribute key>'] == "<a certain value>":
return "gray"
else:
return "#03A9F4"
serve(text, entities, color_map_func=color_map_func)Note that the color_map_func has higher priority than color_attr_key. When provided, the color_attr_key will be overwritten.