Phil Resnik's ACL keynote: A new balancing act
« previous post | next post »
On July 5, Phil Resnik delivered a keynote address at the 64th Annual Meeting of the Association for Computational Linguistics, with the title "A New Balancing Act: Reflections on the Relationship between Computational Linguistics and AI":
In this talk I argued that the field of computational linguistics – a term that includes NLP as its engineering-research subdiscipline – is experiencing a “success catastrophe”. The commercial success of LLM-based AI has thrown three key aspects of our research community out of balance. Here are three balancing acts we face:
First: Like any research community, we can recognize and take advantage of the knowledge obtained in earlier generations of work; we can also lean into new approaches.
Second: We can focus on language as language, which is to say, the properties of language that make it distinctive and human; we can also treat language as an input/output modality for AI systems.
Third: We can emphasize our role as a research community, where our primary purpose is to contribute to the stock of human knowledge; we can also emphasize our role in making sure that our young members have a path forward to get jobs – particularly jobs in industry since the path into academia is never a sure bet and for many of them industry is the goal.
In each of these pairings, the central importance of the former has given way to the overwhelming dominance of the latter.
Moreover, because of the concentration of power and the convergence on methods that require enormous resources, our natural corrective processes – the historical pendulum-swings back and forth – are fundamentally broken. We have lost scientific pluralism.
These losses of balance are bad for the research enterprise, where the explosion of “let me demonstrate my skills to use LLMs and move a score” submissions has brought quality control via peer-review to its knees.
These losses of balance are bad for society, where deployment of systems without a solid understanding of their underlying theories has already caused significant harms to people.
And, I argue, these losses of balance represent an existential threat to our community. AI and LLMs are now inextricably a part of what we do. But if that’s all we do, if we abandon our core value of contributing to human scientific and engineering knowledge about language as language, then we lose who we are. We become just another entry on the list of machine learning conferences.
Phil calls this a "new" balancing act, in reference to a workshop and book from 30 years ago, published as "The Balancing Act: Combining Symbolic and Statistical Approaches to Language", Judith Klavans & Philip Resnick, eds.
Metaphors involving balancing and pendulums can be found in both the old and the new "balancing act" texts (read and hear the whole new thing — the old one is dying on library shelves and in used book stores…).
The intended interpretations of these balancing/pendulum metaphors are certainly valid and useful, but in "The Future of Computational Linguistics: On Beyond Alchemy" (2021), Ken Church and I sketched four socio-scientific eras, and suggested a different metaphor, namely seasonal migration of herbivorous herds:
This description suggests a winner-take-all picture of the field. In fact, the field has always benefited from a give-and-take of interdisciplinary ideas, making room for various combinations of methodologies and philosophies, in different proportions at different times. Logic played a larger role when rationalism was in fashion, and probability played a larger role when empiricism was in fashion, and both logic and probability faded into the background as deep nets gave procedural life to an associationist (rather than statistical) flavor of empiricism. But at every stage, there have been research communities of various sizes inhabiting or exploring different regions of this dynamic landscape, motivated by their various ideological visions, their preferred methodological tools, and their substantive goals. The decades have seen various different communities prosper, decline almost to extinction, and then grow again, waxing and waning in different rhythms. The seeds of the next dominant fashion can always be seen in research communities that seem marginal at a given stage.
Stephen Goranson said,
July 23, 2026 @ 11:49 am
Others will know better: is AI, for Linguistics, better seen as a current enthusiasm and/or a game-changer?
Mark Liberman said,
July 23, 2026 @ 12:52 pm
@Stephen Goranson: "is AI, for Linguistics, better seen as a current enthusiasm and/or a game-changer?"
Why not both?
The same question was asked about the statistical modeling version of machine learning back 30-40 years ago, and the same answer was appropriate.
And ditto for the formal logic version of AI 40-50 years ago, again with the same answer.
And again likewise for the role of information theory a few decades before that — see the discussion in my 2010 obit for Fred Jelinek.
Olaf Zimmermann said,
July 23, 2026 @ 4:01 pm
I thought that LLMs were yet another offshoot from NLP research, and an extremely hardware-dependent one at that. Am I missing something?
Mark Liberman said,
July 23, 2026 @ 4:37 pm
@Olaf Zimmermann: "I thought that LLMs were yet another offshoot from NLP research, and an extremely hardware-dependent one at that. Am I missing something?"
Yes. LLMs (and other "deep nets") have roots in neuro-psychology, physics, coding theory, cryptology, and other fields — see here and here and here for a (starter) sample of sources.
Olaf Zimmermann said,
July 23, 2026 @ 6:10 pm
@Mark Liberman
Thank you for the 'readings' – they were, are, contributaries to NLP, in my humble opinion. LLMs are to me what detergents became in the late '60s: "New and improved! With enzymes!" They are therefore of interest to sociologists studying the 'will to believe', but contribute little to our understanding of language. (Unprincipled parameters and morphemising tokenisers have their limitations.)
I may, of course, as always, be entirely mistaken.
JPL said,
July 23, 2026 @ 11:59 pm
@Olaf Z: "… but contribute little to our understanding of language."
P Resnick, in the OP: "… properties of language that make it distinctive and human;"
As far as research programs are concerned, I don't see why the practical successes of AI should have anything to do with what computational linguistics takes as its aims. LLM- based AI is all about engineering solutions for automatically producing interpretable texts, and its models have achieved an accurate statistical analysis of the collocation relations in texts; pure inquiry into the phenomenon of language is all about the human significance of language, which is best identified as "human thought using language", to focus its central role. A language system's function is not just to express thought, as Chomsky always says, but to enable thought to function more effectively (than it does for animals that don't have language) as a way (a tool, if you will) to understand a complex and problematic world. How can you puzzle about what language is without including its role wrt thought? The text doesn't connect to the world; it's the thought that connects to the world. AI tells us nothing about, and is not interested in, how that works. (Odd that Chomsky has often said that the function of language is to express thought, but his approach, in focusing only on logical "well-formedness" or grammaticality, shows no interest in the thought language is there to express and enable about the world.)