Procter & Gamble

Pawel Bober

In 2022, getting your electricity bill inquiry resolved by a bot in no time or having your Teams meeting transcribed in real time hardly surprises anyone. A (perfectly entertaining) Guardian article authored by a GPT-3 robot, or a computer video Asteroids game developed by an AI system Codex still raise a few eyebrows. A branch of AI - Natural Language Processing, in short NLP, is transforming the world as we know it but the best is yet to come. Enterprises can tap into NLP to unlock a significant share of AI driven direct revenue gain potential.

Right time, shocking results

The story of NLP, like every good show, has featured an impressive cast, offered countless twists and turns, and after the last chapter set in the pandemic, left us breathless, but waiting to see more. In its close to a hundred-year history, NLP came a long way to deliver shocking outcomes.

The initial attempts like translating machines from the 1930s or the automated Russian to English text translation program that Georgetown University and IBM launched in 1954 did not live up to the expectations. Alan Turing contributed to the NLP field his human imitation test for AI, Noam Chomsky the “Universal Grammar” with codified rules and Roger Shank the concept of tokens that help better grasp the meaning of a sentence. But it was the application of Machine Learning and Deep Learning that step changed the NLP outcomes and finally freed us up from specifying language rules.

Since 2015, valuations of AI startups soared and when the COVID-19 pandemic hit, NLP was mature enough to change the enterprise game. Social and economic disruptions massively catalyzed the adoption of NLP. Enterprises facing lockdowns doubled down on automation and customer self-service. Chatbots entered the mainstream (WHO and CDC also launched theirs). NLP has been used to extract new insights from COVID biomedical research and to battle misinformation in social media.

Results matched the shocking reality. Companies reported 6.3 percent of direct revenue gains on average directly attributable to AI. And the NLP powered virtual agent technology with 3 percent revenue gains was the single biggest contributor to that result – pointed out IBM’s mid-pandemic study The Business Value of AI.NLP budgets soared. In the 2021 NLP Industry Survey, 60 percent of tech leaders said that their NLP budgets grew by at least 10 percent, and33 percent pointed out an increase in excess of 30 percent compared to the previous year.

Driving the transformation

NLP became integral to the enterprise but tech leaders continue to be challenged by its time to market, talent availability, business context understanding and ethics. Proven industry best practices can help.

1)Right bets. Successful leaders prioritize NLP use cases according to their benefit, applicability to own value chains, untapped potential, production readiness and scalability. Speech recognition, chatbots, virtual assistants, emotion AI and large pretrained language models use cases are generally considered to have transformative enterprise benefits. Speech synthesis, text classification, semantic search, text summarization use cases are generally attributed with moderate benefits – Gartner points out.

The actual benefits will depend on the industry sector and specific enterprise value chain. The overwhelming part of enterprise investments in NLP should support use cases that can be productionalized and scaled up. The remaining budget can strategically target NLP innovation opportunities.

2)Platforms. The effective use of technology platforms can help shorten the time to market and boost resource productivity. For example, training a massive language model should be done centrally to enable rapid reuse because of its sheer size – Hugging Face estimated that it is going to take about 4 months to train them BigScience language model with 176B parameters on 384 A100 GPUs. Digital enterprise IT platforms that provide cloud, security, data and analytics platforms (including MLOps toolset) provide capabilities widely reusable across the enterprise. Business enablement platforms with integrated less broadly reusable vertical NLP capabilities (e.g. maintenance logbook processing, contract analysis) in turn support individual NLP applications. For relatively mature, commodity areas consider licensing and re-channeling your resources to integration and innovation. Hyperscale cloud providers, once they decide to enter a market and have enough time to iterate, will typically offer a reasonably priced and feature-rich platform products e.g. Microsoft and Google are both Forester’s 2021 leaders in Cognitive Search.

3)Ethics. Infuse your NLP initiatives with ethics from the start. The story of Amazon who had to abandon their automated resume screening tool, as described by Reuters 2018 article “Amazon Scraps Secret AI Recruiting Tool that Showed Bias Against Women” and, the story of Microsoft Tai, a chatbot which was shut down due to its offensive comments, illustrate the lasting negative effects of unsuccessful NLP initiatives. The emerging AI legislation like the EU’s AI Act or UK’s ICO AI Auditing Framework provide additional motivation for structured Responsible AI approaches. The five most common sources of bias in NLP are data selection, annotations, representations, models and research design. However, methodologies and algorithms exist to mitigate them. Also, as Stanford’s Ethical and Social Issues in Natural Language Processing course points out, NLP has a dual nature – you can use it in a way that does not harm but also it can help you do good e.g. to detect bias, hate speech, abuse, fake news and political misinformation.

What’s next

We will see more. Considering further adoption and expected new NLP advancements we should expect further increase in NLP revenue gains for the enterprises. Enterprises moving their AI initiatives from pilots to optimization phases reported an incremental revenue gain of 7.2 percent according to The Business Value of AI study from IBM. Larger language models like GPT-4 expected in the summer 2022, domain specific training sets and hybrid human-digital workforce will open up a new chapter in the NLP history.