eWEEK content and product recommendations are editorially independent. We may make money when you click on links to our partners. Learn More.
Natural language technologies are quickly gaining traction in the enterprise. We are seeing huge numbers of companies every week rolling out text analytics solutions, speech recognition systems, chatbots and every other language-processing use case imaginable.
It’s no surprise that over the past three years natural language processing (NLP) has become one of the most dominant domains in data science. NLP also is an umbrella term for subfields such as natural language understanding (NLU), natural language generation (NLG) and natural language interaction (NLI).
With companies such as Google and Microsoft producing new discoveries on a consistent basis, NLP has made quantum leaps in terms of accuracy, speed and methodology to aid computer scientists as they tackle complex issues. Today, NLP is one of the most researched fields in artificial intelligence (AI).
In this eWEEK Data Points article, Dillon Erb, co-Founder and CEO of Paperspace, describes five important natural language processing trends impacting the enterprise in 2021 and helps distinguish between trends that have legitimate promise and those that are overhyped.
Further reading
Data Point No. 1: True NLU has a long way to go.
Human language is complex — language is a proxy for human thought after all. That makes NLU (natural language understanding) one of the so-called “AI-hard” problems — because to “solve” NLU is to “solve” generalized AI.
But as a subtopic of AI, NLU gets a lot of attention because the possible applications are so exciting and there are so many enabling technologies that are showing real value in the wild already as partial NLU solutions.
So while the first data point is that complete NLU is not close right now, it’s also true that NLU applications are already proving themselves in the enterprise.
NLU is used to perform sentiment analysis on customer help requests and to understand questions posed to digital assistants like Siri and Alexa. It also translates text between languages in multilingual neural machine translation services, like Google Translate.
NLU is all around us, but we must accept that there will still be years of partial NLU before there is full generalized intelligence on the web.
Data Point No. 2: Models are improving rapidly, and this is inspiring enterprises to prepare.
Arguably the most famous AI model in the world right now is GPT from OpenAI. The latest version GPT-3 was released earlier this year.
In terms of GPT’s progress, GPT-2 launched in February 2019 and made a significant impact because it was trained on 1.5B parameters. GPT-3 launched nearly 18 months later and pre-trained on 175B parameters–an increase of two orders of magnitude. These releases instantly generate worldwide media attention because of the implications for fake news generation, generative art, codebases that write themselves and more.
The progress of language models like GPT is inspiring many new approaches and applications for machine intelligence in the enterprise–from being able to write web apps by describing them in English to mimicking language patterns of public figures to training on medical literature to provide diagnoses.
Data Point No. 3: Narrow use cases are bringing success.
Hundreds of thousands of people do GPU computing, machine learning (ML), and deep learning–especially at the enterprise level. One thing that comes up repeatedly is that projects must be narrowly scoped in order to achieve success and drive business value.
Many of the AI projects like OpenAI’s GPT-3 that attract significant media attention are doing so because of the inspiring visions that these technologies enable for our collective future. But the enterprise projects that are driving real value today are narrowly scoped, narrowly deployed and drive very specific business value.
For example, when Adobe Photoshop ships a feature that automatically selects the subject of an image from the background, this is a narrow use of computer vision that immediately provides real value to the end user. Photoshop users don’t have to spend hours knocking out backgrounds with painstaking pixel-by-pixel precision. This is a perfect example of narrow focus leading to definitive success.
Data Point No. 4: Model maintenance
Unlike traditional software, ML models in production require maintenance, since they can and do degrade in performance over time. This fact combined with the complicated makeup of a production ML model (a model is composed of code, data, computation and performance metrics) has resulted in a boom of CI/CD and MLOps tooling and infrastructure abstraction products.
One of the most exciting things about NLU or NLP (or really any ML application) is that it is possible to improve the model over time. In many applications, a nominal improvement in accuracy of a model by a few percentage points can yield millions of dollars of value to the underlying business. Specific examples include applications such as recommendation engines, financial modeling applications and others.
It’s therefore important to have predictable deployment systems in place to continue delivering value over time.
Data Point No. 5: Treating ML as a software discipline
One of the most valuable ways an enterprise can prepare to drive business value from ML is to establish ML as a software discipline within the business. Since the vast majority of ML projects get stuck as research experiments run locally on developer or data scientist laptops, it’s important to prioritize systems that help your company deploy models to production and deliver value.
Treating ML as a software discipline means having an established pipeline to take projects from research into production. It means giving ML the same DevOps attention and resources as the rest of your software stack. It also means versioning code, data, and ML models themselves. It’s also critical to streamline infrastructure provisioning so that anyone within the organization is empowered to deploy a model at scale.
The end goal is to make AI a strength within the organization. In the realm of ML, it makes sense to have strong processes in place so that value can be created as soon as possible. The most profitable models get more profitable over time — so starting now is essential. There is no time to wait.
If you have a suggestion for an eWEEK Data Points article, email cpreimesberger@eweek.com.