Yes, I’m many fields ML has been exaggerated, but it’s just a matter of time. Most companies aren’t using ML effectively yet.
But I know for a fact when you improve NLP, you _can_ improve almost all white collar jobs. NLP the past two years has been crazy, and we haven’t seen anything yet.
For e-commerce search, a common problem is how to give relevant results for misspelled queries or unique queries. Text classification and/or NER can often identify the product type, brand, etc and hopefully give a relevant result which leads to a sale.
Another example is data entry or business analysis roles. Often they will manually extract data from sources and put it into Excel or some other BI tool. NLP ML has gotten good enough in just the last couple years that much of the data extraction can be automated, but you still need a human to verify accuracy.
* NER (name-entity recognition) is becoming insanely good (99+%), this can help with everything from data security to automated form filling, to automated medical billing, etc. Multiple billion dollar industries will be shook up by this.
* My own work on data generation is pointing to reduced surface area of security purposes and improved access to quasi-data for data science and application testing[1]
* Translation from images to text descriptions are also a major thing, which when combined with other systems can do things like make suggested diagnosis.
* Text generation systems are impacting how consumers interact. The chat bots these days are getting to the point where it's very difficult to identify if it's a human or not. I can't go into too many details, but see[2]
* There's also my side business which is a search engine for people who would know an answer to a questions: https://insideropinion.com/
In general, NLP is probably going to have a larger impact that driverless cars on our day-to-day lives. We're seeing insanely good story generation (e.g. GPT-2), chat bots, billing, etc. I also think WAY WAY more is yet to come. The transformer (which is spuring most of this) is only a couple years old. We'll see lots more applications as time goes on.
[1] https://medium.com/capital-one-tech/why-you-dont-necessarily...
That in part is why I don’t buy any of the “AI winter” chatter. We’ve previously never had machines powerful enough to run most useful ML nor enough data to train effectively. Today, as in right now, we have both. Things that were neat theories in the 90s are profitable businesses today.
Nobody in the ML field thinks strong AI is around the corner, or even 10-20 years from now. But that misses the point entirely. ML is useable and useful right now, and that’s not going to revert, no matter how many programmers grumble about “AI hype.”