Prelude

Natural Language Processing (NLP) is the branch of artificial intelligence that lets machines read, interpret and generate human language, whether it arrives as typed text, spoken words, images or even gestures. It powers the voice assistant on your phone, the chatbot on a bank’s website, the spam filter in your inbox and the translation tool you use when travelling.

Research papers show what is possible; patents show what companies are betting money on protecting. That makes a patent landscape a useful lens for understanding where an industry is heading, who is leading it and which corners remain open. This three-part series draws on a landscape study of NLP patents published between June 2017 and June 2019.

Examples of NLP in everyday life

 

1. Voice assistants

Siri, Alexa, and voice-enabled ChatGPT use NLP to understand what you say and respond.

 

2. Language translation

Tools such as Google Translate convert text from one language to another.

 

3. Spam detection

Email services analyze messages to identify spam and suspicious content.

 

4. Chatbots

ChatGPT uses language-processing techniques to understand questions and generate human-like responses.

How does NLP work?

NLP commonly involves these steps:

1. Input: The computer receives text or speech.
2. Preprocessing: The text may be cleaned and divided into smaller units called tokens.
3. Analysis: The system identifies meaning, grammar, context, or intent.
4. Output: The system produces a response, translation, classification, or other result.

Example

Important concepts in NLP

Concept	Meaning
Tokenization	Splitting text into smaller units
Sentiment analysis	Identifying emotions or opinions in text
Named Entity Recognition (NER)	Identifying names, places, dates, and organizations
Machine translation	Translating between languages
Text generation	Producing new text using AI

NLP vs. AI vs. Machine Learning

• AI (Artificial Intelligence): The broad field of creating systems that perform tasks associated with intelligence.
• Machine Learning (ML): A branch of AI in which systems learn patterns from data.
• NLP (Natural Language Processing): The field focused on processing human language, using techniques from AI and machine learning.

Scope and Methodology

The data set

The study extracted 1,543 patent records filed or published between 1 June 2017 and 1 June 2019. After screening, 1,183 were judged relevant. The two-year window was chosen deliberately to capture the latest innovation at the time, a period in which neural approaches to language were advancing quickly.

What counted as “relevant”

A reference was kept only if it met two conditions:

• It claimed natural language processing, whether rule-based, machine-learning-based, or a hybrid of both.
• It used speech, text, image or gesture as the input for language processing.

Preliminary screening relied on the title, abstract and claims. The same three elements were used to assign each patent a technology category, while application categories were assigned through a semi-automatic process. To avoid counting the same invention several times, the analysis uses one representative member per patent family. A patent family is the group of filings in different countries that protect the same underlying invention.

What the report covers

• Bibliographic and technology trends, including legal status, geography and filing years.
• Top assignees (the companies and organisations that own the patents) and the strength of their portfolios.
• Prominent universities and research institutes, and their collaborations with industry.
• Key technology insights, opportunities and a brief market overview.

A Taxonomy of NLP Technology

To make sense of more than a thousand patents, the study classified each one along four dimensions: what goes in, what the system is for, how it works, and where it runs.

Dimension	Categories used in the study
Input	Speech, text, image/video, gesture
Purpose	Speech recognition and analytics (question answering, machine translation, language identification); text analytics (question answering, data filtering, machine translation, data identification/modelling, data classification); sentiment/emotion analysis; web browsing; assistance (mobile assistant, automobile assistant, chatbot, customer support)
Technique	Rule-based; trained machine-learning based (many patents combine both)
Server type	Cloud, local, external

 

Rule-based versus machine learning

Rule-based systems follow patterns and grammar rules written by engineers. They are predictable and easy to audit, but brittle when language gets messy. Machine-learning systems learn patterns from large amounts of example data, which makes them far more flexible but also more data-hungry. The study found that the two are often combined.

Why the input type matters

Text is the simplest input because it arrives already digitised. Speech needs an additional recognition step before language understanding can even begin. Images and gestures add further complexity, since the system must first interpret what it sees. Patents that accept more than one input type tend to be more ambitious.

The Overall Filing Picture

Legal status

The 1,183 relevant patents and publications, 1,158 (about 98%) are alive, meaning still pending or in force, and only 25 are dead. That is expected for a young and fast-moving field where most filings are recent.

Category	Count	Detail
Alive	1,158	872 applications pending; 286 granted
Dead	25	7 expired; 10 rejected, refused or suspended; 8 withdrawn or surrendered
Total relevant	1,183	One member per family

Where the patents are filed

Geography tells you where companies see commercial value and where they expect to enforce their rights. Based on expanded family data, the distribution is heavily weighted toward two jurisdictions.

Jurisdiction	Share of patents/publications
United States (US)	56%
China (CN)	24%
South Korea (KR)	4%
European Patent Office (EP)	3%
Japan (JP) and India (IN)	2%
Others (RU, CA, AU, DE, GB, SE, TW)	3%
International (WO) publications	8%

 

The United States leads with 738 patent families, followed by China with 288. The study also flags Sweden, the United Kingdom, Taiwan and Germany as relatively open territories where research activity is lower and there may be room to build a position.

The filing year matters too. 421 of the relevant patents were published in 2017, the largest single-year share in the window. Of those, 161 were US filings describing artificial intelligence for language processing, a sign that machine learning was already the centre of gravity.