Do people ask good questions?

Rothe, Anselm; Lake, Brenden M.; Gureckis, Todd M.

doi:10.31234/osf.io/eanku

Cited by 4 publications

(5 citation statements)

References 16 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…A recent flurry of work has focused on integrating vision and language, leading to creative combinations of computer vision and NLP models. Active research areas include image-caption generation (Chen et al, 2015; Vinyals et al, 2014; Xu et al, 2015), visual question answering (Agrawal et al, 2017; Das et al, 2018; Johnson, Hariharan, van Der Maaten, Fei-Fei, et al, 2017), visual question asking (Mostafazadeh et al, 2016; Rothe et al, 2017; Wang & Lake, 2021), zero-shot visual category learning (Lazaridou et al, 2015; Xian et al, 2017), and instruction following (Hill, Lampinen, et al, 2020; Ruis et al, 2020). The multimodal nature of these tasks grounds the word representations acquired by these models, as we discuss below.…”

Section: Desideratamentioning

confidence: 99%

Word meaning in minds and machines.

Lake¹,

Murphy²

2023

Psychological Review

Self Cite

View full text Add to dashboard Cite

Machines have achieved a broad and growing set of linguistic competencies, thanks to recent progress in Natural Language Processing (NLP). Psychologists have shown increasing interest in such models, comparing their output to psychological judgments such as similarity, association, priming, and comprehension, raising the question of whether the models could serve as psychological theories. In this article, we compare how humans and machines represent the meaning of words. We argue that contemporary NLP systems are fairly successful models of human word similarity, but they fall short in many other respects. Current models are too strongly linked to the text-based patterns in large corpora, and too weakly linked to the desires, goals, and beliefs that people express through words. Word meanings must also be grounded in perception and action and be capable of flexible combinations in ways that current systems are not. We discuss promising approaches to grounding NLP systems and argue that they will be more successful, with a more human-like, conceptual basis for word meaning.

show abstract

Section: Desideratamentioning

confidence: 99%

Word meaning in minds and machines.

Lake¹,

Murphy²

2023

Psychological Review

Self Cite

View full text Add to dashboard Cite

show abstract

“…Following a growing body of work in compositional Bayesian models ( 38 , 58 , 59 , 61 , 73 – 88 ), we assume that the representations learners must discover are built by combining primitives in a language of thought (LOT) ( 49 ) to form the mental analog of programs. In this setup, learners observe data (here, strings) and compare hypotheses that are built out of primitives, as a way to explain the data, much as scientists might consider possible physical laws which are compositions of mathematical operations (e.g.,

).…”

Section: Formal Modelmentioning

confidence: 99%

One model for the learning of language

Yang

Piantadosi

2022

Proc. Natl. Acad. Sci. U.S.A.

View full text Add to dashboard Cite

A major goal of linguistics and cognitive science is to understand what class of learning systems can acquire natural language. Until recently, the computational requirements of language have been used to argue that learning is impossible without a highly constrained hypothesis space. Here, we describe a learning system that is maximally unconstrained, operating over the space of all computations, and is able to acquire many of the key structures present in natural language from positive evidence alone. We demonstrate this by providing the same learning model with data from 74 distinct formal languages which have been argued to capture key features of language, have been studied in experimental work, or come from an interesting complexity class. The model is able to successfully induce the latent system generating the observed strings from small amounts of evidence in almost all cases, including for regular (e.g., an, (ab)n, and {a,b}+), context-free (e.g., anbn, anbn+m, and xxR), and context-sensitive (e.g., anbncn, anbmcndm, and xx) languages, as well as for many languages studied in learning experiments. These results show that relatively small amounts of positive evidence can support learning of rich classes of generative computations over structures. The model provides an idealized learning setup upon which additional cognitive constraints and biases can be formalized.

show abstract

“…More recent work has shown how learners might construct generative mental theories of entire structures like the integers (infinite, ordered, and discrete) from a simpler basis that does not presuppose this conceptual structure, but is able to acquire many different structures across domains (Piantadosi, 2021). This general approach of learning procedures and representations is notable in drawing on inferential processes that have been argued for independently in concept learning (Amalric et al, 2017; Calvo & Symons, 2014; Depeweg et al, 2018; Erdogan et al, 2015; Goodman et al, 2008; Goodman et al, 2015; Lake et al, 2017; Piantadosi & Jacobs, 2016; Romano et al, 2018; Rothe et al, 2017; Rothe et al, 2016; Wang et al, 2019; Yildirim & Jacobs, 2015), potentially showing how number systems may be constructed like other—even artificial—systems of rules that adults acquire and fluidly manipulate.…”

Section: Symbolic Number Learning Is a Difficult Developmental Processmentioning

confidence: 99%

The cultural origins of symbolic number.

O’Shaughnessy¹,

Gibson²,

Piantadosi³

2022

Psychological Review

View full text Add to dashboard Cite

In order to explain the unacceptability of certain long-distance dependencies --termed syntactic islands by Ross (1967) --syntacticians proposed constraints on long-distance dependencies which are universal and purely syntactic and thus not dependent on the meaning of the construction (Chomsky 1977(Chomsky , 2006. This predicts that these constraints should hold across constructions and languages. In this paper, we investigate the "subject island" constraint across constructions in English and French, a constraint that blocks extraction out of subjects. In particular, we compare extraction out of nominal subjects with extraction out of nominal objects, in relative clauses and wh-questions, using similar materials across constructions and languages. Contrary to the syntactic accounts, we find that unacceptable extractions from subjects involve (a) extraction in wh-questions (in both languages); or (b) preposition stranding (in English). But the extraction of a whole prepositional phrase from subjects in a relative clause, in both languages, is as good or better than a similar extraction from objects. Following Erteschik-Shir (1973) and Kuno (1987) among others, we propose a theory that takes into account the discourse status of the extracted element in the construction at hand: the extracted element is a focus (corresponding to new information) in wh-questions, but not in relative clauses. The focus status conflicts with the non-focal status of a subject (usually given or discourse-old). These results suggest that most previous discussions of islands may rely on the wrong premise that all extraction types behave alike. Once different extraction types are recognized as different constructions (

show abstract

Do people ask good questions?

Cited by 4 publications

References 16 publications

Word meaning in minds and machines.

Word meaning in minds and machines.

One model for the learning of language

The cultural origins of symbolic number.

Contact Info

Product

Resources

About