Abstract
The impressive performance of the latest large language model (LLM), ChatGPT, has been currently in the news and has received a lot of very positive and very negative hype. Current LLMs generate coherent, well-formed, and relevant text in response to prompts by the human user. ChatGPT also seems to operate in a more dialogical and interactive way, which makes its usability much more intuitive. These successes raise the possibility that these models are on their way to becoming thinking agents capable of executing tasks that require common sense knowledge, abstract reasoning, and rationality.
Some linguists, computer scientists, and psychologists have been raising doubts about the feasibility and desirability of such models, including the famous recent paper “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” by Bender and colleagues (aka “the parrots paper”; 2021). In various formal and informal publications (e.g., tweets, Facebook posts), researchers raise concerns about the biases in data that are perpetuated and amplified in the output of such models, disempowerment of communities with no access to such expensive technology, and serious environmental costs. In the paper, "Climbing towards NLU: On meaning, form, and understanding in the age of data" (aka “the octopus paper”; 2020), Bender and Koller also the raise the issue that it is in principle impossible that such models will be able to capture the complexity and nuance of human languages. The octopus paper is a position paper that summarises why linguists should be pessimistic about the prospects of capturing human language abilities by means of training LLMs solely on text. The main argument is that since LLMs are primarily based on pattern recognition and statistical associations, rather than embodied interaction with the world and referential semantic abilities, they are bound to fail as useful or even convincing communicative partners for humans, while incurring the costs pointed out in the parrots paper. The main argument in the octopus paper relies on the presumed—and taken for granted—distinction between form and meaning that prevents machines trained solely on form to acquire genuine understanding that relies on meaning.
Both of us have found the arguments presented for this view from somewhat lacking to totally unconvincing but, we have discovered, for rather different reasons. We suspect that this has to do with our respective distinct theoretical orientations and views about language and the human mind. For this reason, we would like to discuss these issues in a public forum with feedback from other researchers to see whether this is the case and whether there is a possibility of adopting some kind of conciliatory perspective.
References
Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021, March). On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?🦜. In Proceedings of the 2021 ACM conference on fairness, accountability, and transparency (pp. 610-623). doi: https://doi.org/10.1145/3442188.3445922
Bender, E. M., & Koller, A. (2020, July). Climbing towards NLU: On meaning, form, and understanding in the age of data. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (pp. 5185-5198). doi: http://dx.doi.org/10.18653/v1/2020.acl-main.463
The Linguistic Forum
The Linguistic Forum is an informal meeting place for all linguists working or studying at the Faculty of Humanities. The Forum is a new faculty-wide seminar series that has financial support from the Faculty. Our aim is to promote the sharing of knowledge and collaborations between the Faculty’s linguists.
Mailing list
Register to our mailing list https://medarbetarportalen.gu.se/internt-sprak/formular/sprakvetenskapligt-forum/ to get information on upcoming events.
