Skip to content
Back to search results

Information Extraction — Turning Text into Structured Records

Type
NLP structured-information task
Font size

What does it do?

Information extraction (IE) is a natural-language-processing task that converts unstructured text into structured information — such as events, relations, and attributes — so that facts buried in free-form documents become machine-readable records.

The task was consolidated in early surveys by Cowie and Lehnert (1996) and later by Grishman (2012).

The intuition behind the method

A document tells a story in free-form language, but a database needs neat rows and fields. Information extraction reads through the text the way an analyst would highlight the who, what, and when, then writes those highlights into a fixed schema. Instead of leaving the facts trapped in prose, the method pulls out the entities and the relations between them and lays them out as structured records you can query.

History

Information extraction was framed as a distinct NLP task in the surveys of Cowie and Lehnert (1996), who consolidated the goal of pulling structured facts out of free text, and was later revisited by Grishman (2012) in the Handbook of Natural Language Processing. The task has grown from hand-crafted pattern rules toward statistical and learned models for finding entities, relations, and events.

When to use it

Information extraction fits when you have text data and need to turn it into structured records of events, relations, or attributes, and when you can define a domain-specific ontology describing what to extract. A reasonable volume of text helps — the method expects on the order of at least fifty documents. Without text data, or without an agreed schema for what counts as a fact, IE cannot run.

Limitations and considerations

  • Requires a clearly defined domain-specific ontology before extraction can begin.
  • Needs a sufficient amount of text data; very small corpora give too little signal.
  • Extraction quality depends heavily on how well the schema and patterns match the language of the corpus.

Source conditions do not replace independent scientific review or validation by an analysis engine.

Strengths

  • Turns unstructured text into structured, queryable records of events, relations, and attributes.
  • Schema-driven: the ontology lets the same pipeline be retargeted to different domains.
  • Combines rule-based pattern matching and trained models, so it adapts to the available data and resources.

Applications

  • Building structured knowledge bases from large document collections.
  • Extracting events and their participants from news or report archives.
  • Capturing relations and attributes from domain text in health, social science, economics, and business settings.
  • Populating databases with facts mined from unstructured documents.

How it works

How it works

1

A domain-specific ontology or schema is defined first: the entity types, relations, events, and attributes that the system should look for. Because IE only extracts what the schema describes, this step decides what counts as a fact worth capturing in a given domain.

2

The raw text is cleaned and segmented into analysable units — sentences and tokens — so that the extraction patterns and models operate on consistent input. Without this normalisation the downstream matching becomes unreliable.

3

Entities, relations, events, and attributes are located in the text. This is done with pattern-matching rules over linguistic structure, with trained models, or with a combination of both, each spotting the elements defined in the ontology.

4

The extracted elements are assembled into structured records that conform to the schema — filling the slots of an event, linking the participants of a relation, or attaching attributes to an entity. The result is machine-readable structured information drawn from the original unstructured text.

Common pitfalls

  • Skipping ontology design — without a defined schema the system has no target and extracts inconsistent or meaningless records.
  • Applying patterns or models built for one domain to another degrades extraction accuracy.
  • Neglecting preprocessing leaves noisy tokens and broken sentence boundaries that derail downstream matching.
  • Running on too little text (well under fifty documents) yields too few examples for the patterns or models to work reliably.

Frequently asked questions

How is information extraction different from named-entity recognition?

Named-entity recognition only locates and labels entities such as people, places, and organisations. Information extraction is broader: it also identifies the relations between those entities, the events they participate in, and their attributes, and assembles all of this into structured records that follow a schema.

Why do I need an ontology before running it?

The ontology, or schema, defines exactly what counts as a fact in your domain — which entity types, relations, events, and attributes to look for. Information extraction only captures what the schema describes, so without one the system has no target and produces inconsistent output.

Do I need labelled training data?

It depends on the approach. Pattern- or rule-based extraction operates over linguistic structure and can work without labelled examples, while trained models learn from annotated data. Many systems combine both, so the data you need varies with the method you choose.

How much text do I need?

Information extraction expects a reasonable volume of text — on the order of at least fifty documents. With far fewer, the patterns and models have too few examples to extract reliably.

Related methods

Sources

  1. Cowie, J. & Lehnert, W. (1996). Information Extraction. Communications of the ACM.

    Open DOI reference
    Copied.
  2. Grishman, R. (2012). Information Extraction. In Handbook of Natural Language Processing.

    ISBN: 9781420085921
    Copied.