Thoughts on NLP 30 years ago

 About 30 years ago I gave a talk to my colleagues at a big Dictionary/ELT publisher. I can't remember the exact title, but it was something like 'Why you need to find out about NLP'. That kind of thing.

This is a little techy, but only a bit, and I hope you may find it interesting. The NLP I was referring to was not the 'Neuro-Linguistic Programming' that is sometimes mentioned after that acronym ("Every day in every way I am getting better and better", that sort of thing). It is the other NLP, which stands for Natural Language Processing. 

Basically, this NLP is everything that you can do with computers that relates to not their own languages of Fortran, Basic, etc etc (you can hear me spitballing here I know) but rather to the computational processing, manipulation and reproduction of our own 'natural' languages such as English, Chinese, Amharic, whatever. 

You may guess that there have been some rather large developments in the last 30 years but, even then, there were things that I could demonstrate that showed the direction of travel. There were, for example, very basic chatbots, where you could ask a text question and it would give you back things that were sort of replies, by processing your input into something that it could, if not understand, at least respond to in a way that seemed partly human. 

There were also the first steps in automatic machine translation, which seemed so far off then that plenty of computer and linguistic types said it would never happen. But if you are in a foreign country now, you will know that for all intents and purposes, functional machine translation exists and is used every day. The apps that we will use in China this summer, such as WeChat and Alipay have got embedded translation functions in them which can, for example, scan a menu and translate it into your language. 

30 years ago we had the first versions of 'grammar checkers' too, which were as intrusive as the modern ones but much less accurate. I remember mine could not see any error in 'I don't can come' because in one sense of the verb can you are allowed that structure (ie we don't can sardines, we only can salmon). Things are so much more sophisticated now, and you may often find that the setup on your machine will just correct your grammar without asking, if you let it. Which brings us onto that other ubiquitous piece of language software in 2026, ChatGTP and its ilk. I could have drafted this entire piece in a couple of minutes had I chosen to let a programme do it. That is so so far beyond what we could have imagined in 1996 that I don't think anyone had even suggested it might be possible.

There were a couple of points I wanted my colleagues to understand clearly, and one was that we as citizens were no longer safe and anonymous in numbers. Say a malevolent government wanted to listen to your phone calls or read your writing in the 1960s, to root our traitors. The only solution for them was a grinding army of readers and listeners to look at every word we had spoken or written. But as soon as automatic programming and searching of large sets of data were becoming feasible, you could write a programme to search for key words such as 'Semtex' or whatever the code for it was, and know who had used it. I just asked AI to find out about early US government language research related to security and the military and this is what it came up with (in 2 seconds)

NLP and the military in the US

One of the reasons that I had such an active interest in this field was that the company I worked for, Longman (now part of Pearson) had been very successful in the very early steps of selling structured language data to this new NLP market. One of our dictionaries used a very controlled defining language, which gave a better chance to computers of extracting meaning and semantic links from the data. Big software companies were also early customers. In fact I also managed a project seeking to extend the usefulness of our data by adding in something called 'complementation patterns' for 500 common verbs. Here's how it works. You take a verb like 'elect' and you show all the ways that it can be used and the typical things it refers to. So for elect you could:

elect somebody (eg to elect a new chair of the board)

elect somebody something (where the 'something' is a role or post) (eg to elect her president)

elect to do something (eg he elected not to inform us)

This is a simple example and you can imagine how much more complex it becomes with a very like stand or run.

We worked on this for about a year, and I can remember near the end of it realising that I was listening to the radio news in the morning and parsing their sentences live. Time to step back from it. 

So if I was doing that talk now, what would I say? What are the biggest achievements, threats, possibilities for the future. Well I think that the Babel Fish from 'The Hitchhikers' Guide to the Universe' is well within sight. I'm not sure I like the idea, but hell, it's convenient. 

I believe that our use of shortcuts to express ourselves using ChatGPT and its ilk will continue to boom and I also fear that at some time we will realise that we have lost something essential in our souls by using it, but hey I could be wrong (inject lighter note here to balance dread, ChatGPT). I worry that our ability to learn new things, to empathise, to tolerate boredom will all be reduced by the quick wins of AI. On the plus side, I hope that the use of NLP and AI may help us to make medical discoveries, avoid fatal errors caused by language misunderstandings, rescue endangered languages, maybe even make us wiser in some way. 

But the main thing I feel when looking back at 30 years of change is a firm belief that, while I may have an idea of the general trends, I could never predict what another 30 years will produce. And maybe I would rather not.