Many people who are retiring like to reminisce about the “good old days” when they were young. I’ve tried to avoid this, but some younger people seem genuinely curious about what NLG and NLP were like in distant past, so I thought I’d write something about what it was like to be an NL researcher in 1990, when I got my PhD.

Incidentally, my PhD was on “Generating Appropriate Natural Language Object Descriptions”. If anyone is curious, the best part of it was work on generating descriptions that referred to objects. This was published as a paper in ACL 1990. I kept on working on this topic after my PhD, collaborating with Robert Dale, and we published a journal article on the topic in 1995. Publishing an expanded and improved version of an ACL paper as a journal article was common and indeed often expected in 1990, although usually the time gap between conference and journal paper was less than 5 years!

Research careers and infrastructure

I think early career researchers (like me) had it easier in 1990 compared to what I see in 2026. People often moved directly from PhD into good faculty positions (I did a post-doc, but many people skipped this career step). There was less pressure to publish large numbers of papers. I had no papers when I started my PhD, and three conference papers when I finished, and this was pretty typical. I think research grants were also easier to get.

On the other hand, in 1990 it was very important to be at an institution which had an excellent research library, because you needed hard copy journals or conference proceedings to read a paper. The Internet in 1990 was mainly a tool for emailing academic colleagues, and browsing USENET forums (which I did a lot of as a PhD student…), not for downloading papers. In 2026, researchers all over the world have access to almost all recent papers in NLP and many other fields, which is absolutely a change for the better.

Research venues

In 1990, the most prestigious NLP venue was Computational Linguistics journal, which published 14 papers in 1990, and also 17 book reviews (books were much more important in 1990). Conference papers were less prestigious (I heard them described as “research McNuggets” in 1990), with ACL being the best conference; ACL-1990 had a grand total of 39 papers (including mine). Incidentally all ACL reviewing was done by 12 senior members of the community. The biggest NLP conference in 1990 was COLING, which had 200 papers. There were also hundreds of papers at more specialised venues, such as INLG (25 papers). So altogether there were roughly 500-600 NLP papers published in 1990. Which meant that an interested researcher could read at least title and abstract of every paper relevant to her (if she had access to hard-copy journals and proceedings!).

In contrast, in 2025 there were well over 10000 papers added to the ACL Anthology, and which means it is impossible for an individual researcher to read title and abstract of all relevant papers. I personally mostly rely on recommendations from colleagues and students (and sometimes LLMs) when deciding what to read, which means I miss interesting and important papers because I am not aware of them.

The community is also less selective in 2025. In 1990, 2.5% of Anthology papers appeared in the top venue of the time, CL journal. In 2025, if we consider ACL and EMNLP full papers, along with CL and TACL journals, as the “prestige” venues, then this amounts to around 3600 papers. Ie, in 1990 2.5% of Anthology papers appeared in the top venue, while in 2025 25% of Anthology papers appeared in the “top” venues.

Research focus

In 1990 the community focused on rule-based techniques and linguistic grammars (comparing unification grammars to tree-adjoining grammars was a big topic in the early 1990s); evaluation was also more linguistic (paper). Data-based techniques were starting to be proposed, and in particular Brown et al suggested using ngram models for statistical machine translation. This was regarded as very speculative idea in 1990, and no one built actual MT systems this way. Indeed, statistical MT did not become better than rule-based MT from a practical perspective until the early 2000s.

I think it says something about the NLP community in 1990 that the Brown et al paper was published in our (very selective) top venue, CL journal, despite being inferior to established approaches at the time. We were interested in “crazy new ideas”, even if they were a decade away from being genuinely useful. In 2026 the NLP research culture is much more hostile to papers which challenge the LLM “orthodoxy” (blog). It is hard to imagine a paper suggesting a radical approach getting published in a good venue in 2026, especially if it showed poor benchmark performance and needed ten years of further research to become competitive.

Commercial focus

From a commercial perspective, there was a lot of excitement in the late 1980s in using NLP to let people query databases in English instead of SQL, and indeed some such products were commercially successful for a while. However, it turned out that people were mainly buying out of curiosity, there was not much genuine need for this kind of thing. DB experts preferred to write SQL directly, people with moderate DB skills preferred GUI-based query builders, and people with no knowledge of databases got frustrated because they would ask English questions which the DB could not answer. In short, this was a cool and impressive-sounding use case (by 1990 standards) which was pretty useless in the real world. I have seen many other cases over the years where AI researchers came up with use cases and demos which sounded impressive but actually were not useful.

Final thoughts

The NLP community in 1990 was much smaller, and worked on topics which sound quaint and out-of-date in 2026; it was also much less commercially successful and was ignored by the mass media. But as a community it had many advantages over 2026, including a more open research culture and less pressure on early career researchers. Being small also meant that we knew each other and what the community as a whole was working on. None of us expected to get rich (I personally passed on that opportunity when I decided to do a PhD instead of working on Wall Street). But it was a great community to be part of, for those of us who loved investigating AI, computation, and language.