The main point of the article could be summarized like so:
We argue that the language modeling task, because it only uses form as training data, cannot in principle lead to learning of meaning. We take the term language model to refer to any system trained only on the task of string prediction, whether it operates over characters, words or sentences, and sequentially or not. We take (linguistic) meaning to be the relation between a linguistic form and communicative intent.
Several NLP papers (cited in the text) use overly confident claims about language models understanding a piece of text or knowledge.
The authors formalize meaning with a set of pairs \(M \subseteq E \times I\), \((e, i)\) representing natural language expressions and their communicative intent. In this framework, understanding means being able to retrieve \(i\) when given \(e\).
They also mention the concept of conventional meaning, i.e. the communicative potential \(s\) of a form \(e\) which is constant across contexts of use of the form \(e\).
When communicating, a speaker has a prior intent \(i\), and has to choose a form \(e\) with the right potential \(s\) that fits \(i\). The argument of the paper is that there isn’t enough signal in a text corpus to learn the above relations in \(M\).
A nice example is the Chinese room experiment, wherein a person is tasked with answering questions in Chinese by consulting a library of Chinese books according to fixed rules. Such a person would look intelligent without having any actual understanding of Chinese.
A nice example is the problem of learning language models on Programming languages. No matter how good a LM is at learning the semantics of the language, it has no way of learning the relations between program inputs and outputs.
The authors provide more example why human-language acquisition is fundamentally different from the flawed language models.
A good LM might learn to give meaningful answers, but will need to store infinite amount of data and will ultimately fail at dealing with the ever evolving set of forms humans use.
Regarding the programming language example, one could argue that by learning the semantics of the language on a large corpus, a LM will eventually have to learn inputs/outputs mapping for intermediate functions in the larger programs, and could build understanding from these small building blocks. However, no explicit mechanisms in current language models allow for combining and re-using building blocks like that.
This is the same with language models, as long as we use neural networks trained with Gradient descent, it will probably never have this power humans have to re-use and create forms to manipulate meaning and deal with a changing world. In a way GPT-3 (Brown et al. 2020) has showed that we can get to a pretty impressive point with sentence prediction alone provided we have enough storage.
Bender, Emily M., and Alexander Koller. 2020. “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, 5185–98. Online: Association for Computational Linguistics.