Named Entity Recognition with SpaCy [with code example]

Named Entity Recognition with SpaCy [with code example]

By the end of this tutorial, you will be able to write a Named Entity Recognition pipeline using SpaCy: it will detect company acquisitions from news headlines.

Demo:

spacy_ner_demo.gif

Learn on practice how to use named entity recognition to mine insights from news in real-time. We will write a Python script to analyze news articles' headlines to learn about all the latest company acquisitions.

In this tutorial, we will:

  • build a data pipeline to fetch real-time news headlines
  • apply pre-trained named entity recognition models provided by Spacy to identify companies that are acquired
  • build an engine that tracks all companies acquisitions



Intro. NLP & Named Entity Recognition

Natural Language Processing (NLP) is a set of techniques that helps analyze human-generated text.

Examples of applying NLP to real-world business problems:

  • chatbots
  • text summarization
  • recommendation engines
  • text autogeneration
  • "trading on news"
  • speech recognition

We try to solve a problem in this article: find all company acquisitions from news articles published online.


Why Named Entity Recognition?

Let's compare these two articles:

"Apple is planning to acquire a news API provider in the near future"

"Apple for breakfast: pros and cons"

Both articles mention "Apple"; however, the first one talks about the company, while the second one is about the fruit.

apple_exampe_ner.jpg

Without named entity recognition, we are not able to even qualify the content.

There are not just organizations (ORG) that the NER can identify. Spacy - Python NLP package that we're going to use - has a pre-built NER pipeline that includes:

  • DATE - Absolute or relative dates or periods
  • PERSON - People, including fictional
  • GPE - Countries, cities, states
  • LOC - Non-GPE locations, mountain ranges, bodies of water
  • MONEY - Monetary values, including unit
  • TIME - Times smaller than a day
  • PRODUCT - Objects, vehicles, foods, etc. (not services)
  • CARDINAL - Numerals that do not fall under another type
  • ORDINAL - "first", "second", etc.
  • QUANTITY - Measurements, as of weight or distance
  • EVENT - Named hurricanes, battles, wars, sports events, etc.
  • FAC - Buildings, airports, highways, bridges, etc.
  • LANGUAGE - Any named language
  • LAW - Named documents made into laws.
  • NORP - Nationalities or religious or political groups
  • PERCENT - Percentage, including "%"
  • WORK_OF_ART - Titles of books, songs, etc.

For example, take a look at the article below after it's processed by the Spacy pipeline:

"On the 15th of September, Tim Cook announced that Apple wants to acquire ABC Group from New York for 1 billion dollars"

spacy_entities_example.jpg


Part I. Named Entity Recognition using Spacy

For now, let's use a pre-defined list of news article headlines to test named entity recognition in Spacy.

In the case of one company that acquires another one, it is fair to assume that there should be at least two ORG tags per headline: at least one for the acquire and at least one for the acquirer.

To install Spacy, run in your console:

pip install spacy

python -m spacy download en_core_web_lg

In your Python interpreter, load the package and pre-trained model:

First, let's run a script to see what entity types were recognized in each headline using the Spacy NER pipeline.

The very first example is the most obvious: one company acquires another one. Spacy NER identified both companies correctly.

2021-10-07 14.31.20.jpg

The third article headline talks about an organization and a person.

2021-10-07 14.31.26.jpg

All headlines talk about some acquisition; however, not all talk about one company acquiring another one.


1. Function to filter only the articles with two ORG entities (and "acquire" in-between)

Also, let's make it a generic function that we can apply to no matter which article.

The function below does two things:

  • detects if a lemma "acquire" is present in the sentence. We want to use lemma - a canonical form of a word - to make sure we cover cases as "acquired", "acquires", etc.
  • check that at least 2 ORG are found

Limitations:

  • can be applied only to one sentence (not on a text)
  • accuracy depends on a NER that you use


Analyzing the dependency using Spacy

Now that we know that a headline is an excellent candidate for an acquisition example, we want to ensure that one company is acquiring another one. For this, we will use dependency parsing with Spacy.

I'm not going to go much into the detail as it is outside of the scope of this article.

Using display, we can visualize the dependency tree.

2021-10-07 14.43.15.jpg

2021-10-07 14.43.21.jpg


2. Function to detect a company(ies) to be bought

The company that is bought should be:

  • of entity type ORG
  • with "head" lemma being "acquire"
  • independency of "attr" or "dobj"

After identifying the token, we:

  • take its noun chunk
  • check for "siblings": other companies that can be bought as well


3. Function to detect an acquiring company(ies)


4. Combine functions and test



Part II. Getting Real-Time Headlines with News API

I'll provide a quick example of how you could get the latest articles that contain lemma "acquire" from 2 sources: prnewswire and businesswire.

We'll use NewsCatcher News API Python SDK for that.

To install News API SDK, run in your console:

pip install newscatcherapi

You can grab your API key here.

We'll make a call to find articles that:

  • contain lemma "acquire" in title
  • published by prnewswire.com or businesswire.com
  • published in the past 24 hours
  • are written in English

Finally, we print out the results.

I leave you here so that you can experiment with actual data yourself.


A few notes on my Spacy NER accuracy with "real world" data

  • Low accuracy with sentences without a proper casing
    1. Low accuracy overall, even with a large model
    2. You'd need to fine-tune your model if you want to use it in production
    3. Overall, there's no open-source high accuracy NER model that you can use out-of-a-box
    By using this website you agree to our Policy.