Gary Marcus wrote: three perspectives from linguists that AI researchers need to know

Everybody knows that large language models like GPT-3 and LaMDA have made tremendous strides, at least in some respects, and powered past many benchmarks, and Cosmo recently described DALL-E but most in the field also agree that something is still missing. A group of engineers at Facebook, for example, wrote in 2019 that:

A growing body of evidence shows that state-of-the-art models learn to exploit spurious statistical patterns in datasets... instead of learning meaning in the flexible and generalizable way that humans do."

Since then, the results on benchmarks have gotten better, but there’s still something missing.

If we had to put our finger on what is still missing, we would focus on these three key elements:

1. Reference: Words and sentence don’t exist in isolation. Language is about a connection between words (or sentence) and the world; the sequences of words that large language models utter lack connection to the external world.

2. Cognitive models: The ultimate goal of a language system should be to update a persisting but dynamic sense of the world. Large language models don’t produce such cognitive models, at least not in a way that anybody has been able to make reliable use of.

3. Compositionality: Complex wholes are (mostly) systematically interpreted in terms of their parts, and how these parts are arranged. Systems like DALL-E face clear challenges when it comes to compositionality. (LLM’s like GPT produce well-formed prose but do not produce interpretable representations of utterances that reflect structured relationships between the parts of those sentences.)

In our view, inadequate attention to these three factors has serious consequences, including:

(a) the tendency of large language models to lose coherence over time, drifting into “empty” language with no clear connection to reality;

(b) the difficulty of large language models in distinguishing truth from falsehoods;

(c) the struggle in these models to avoid perpetuating bias and toxic speech.

 

Now here’s the thing: none of these three elements we have been stressing are news to linguists. In fact, at least since the work of Gottlob Frege in the late 19th century, they have been pretty central to what many linguists worry about. To be sure, none of these three issues has been solved so far; for example, there is still debate about “how much” of our everyday language use actually relies on compositionality, and what the right cognitive models of language should be. But we do think that linguistics has a lot to offer in terms of formulating and thinking about these questions.