Alexander Clark

This interdisciplinary new work explores one of the central theoretical problems in linguistics: learnability. The authors, from different backgrounds---linguistics, philosophy, computer science, psychology and cognitive science-explore the idea that language acquisition proceeds through general purpose learning mechanisms, an approach that is broadly empiricist both methodologically and psychologically. Written by four researchers in the full range of relevant fields: linguistics, psychology, computer science, and cognitive science, the book sheds light on the central problems of learnability and language, and traces their implications for key questions of theoretical linguistics and the study of language acquisition.

The question of whether humans represent grammatical knowledge as a binary condition on membership in a set of well-formed sentences, or as a probabilistic property has been the subject of debate among linguists, psychologists, and cognitive scientists for many decades. Acceptability judgments present a serious problem for both classical binary and probabilistic theories of grammaticality. These judgements are gradient in nature, and so cannot be directly accommodated in a binary formal grammar. However, it is also not possible to simply reduce acceptability to probability. The acceptability of a sentence is not the same as the likelihood of its occurrence, which is, in part, determined by factors like sentence length and lexical frequency. In this paper, we present the results of a set of large-scale experiments using crowd-sourced acceptability judgments that demonstrate gradience to be a pervasive feature in acceptability judgments. We then show how one can predict acceptability judgments on the basis of probability by augmenting probabilistic language models with an acceptability measure. This is a function that normalizes probability values to eliminate the confounding factors of length and lexical frequency. We describe a sequence of modeling experiments with unsupervised language models drawn from state-of-the-art machine learning methods in natural language processing. Several of these models achieve very encouraging levels of accuracy in the acceptability prediction task, as measured by the correlation between the acceptability measure scores and mean human acceptability values. We consider the relevance of these results to the debate on the nature of grammatical competence, and we argue that they support the view that linguistic knowledge can be intrinsically probabilistic.

The question of whether grammaticality is a binary categorical or a gradient property has been the subject of ongoing debate in linguistics and psychology for many years. Linguists have tended to use constructed examples to test speakers’ judgements on specific sorts of constraint violation. We applied machine translation to randomly selected subsets of the British National Corpus (BNC) to generate a large test set which contains well-formed English source sentences, and sentences that exhibit a wide variety of grammatical infelicities. We tested a large number of speakers through (filtered) crowd sourcing, with three distinct modes of classification, one binary and two ordered scales. We found a high degree of correlation in mean judgements for sentences across the three classification tasks. We also did two visual image classification tasks to obtain benchmarks for binary and gradient judgement patterns, respectively. Finally, we did a second crowd source experiment on 100 randomly selected linguistic textbook example sentences. The sentence judgement distributions for individual speakers strongly resemble the gradience benchmark pattern. This evidence suggests that speakers represent grammatical well-formedness as a gradient property.

Learning theory has frequently been applied to language acquisition, but discussion has largely focused on information theoretic problems—in particular on the absence of direct negative evidence. Such arguments typically neglect the probabilistic nature of cognition and learning in general. We argue first that these arguments, and analyses based on them, suffer from a major flaw: they systematically conflate the hypothesis class and the learnable concept class. As a result, they do not allow one to draw significant conclusions about the learner. Second, we claim that the real problem for language learning is the computational complexity of constructing a hypothesis from input data. Studying this problem allows for a more direct approach to the object of study—the language acquisition device—rather than the learnable class of languages, which is epiphenomenal and possibly hard to characterize. The learnability results informed by complexity studies are much more insightful. They strongly suggest that target grammars need to be objective, in the sense that the primitive elements of these grammars are based on objectively definable properties of the language itself. These considerations support the view that language acquisition proceeds primarily through data-driven learning of some form

Indirect negative evidence is clearly an important way for learners to constrain overgeneralisation, and yet a good learning theoretic analysis has yet to be provided for this, whether in a PAC or a probabilistic identification in the limit framework. In this paper we suggest a theoretical analysis of indirect negative evidence that allows the presence of ungrammatical strings in the input and also accounts for the relationship between grammaticality/acceptability and probability. Given independently justified assumptions about lower bounds on the probabilities of grammatical strings, we establish that a limited number of membership queries of some strings can be probabilistically simulated