No abstract

One of the great successes of the application of generalized quantifiers to natural language has been the ability to formulate robust semantic universals. When such a universal is attested, the question arises as to the source of the universal. In this paper, we explore the hypothesis that many semantic universals arise because expressions satisfying the universal are easier to learn than those that do not. While the idea that learnability explains universals is not new, explicit accounts of learning that can make good on this hypothesis are few and far between. We propose a model of learning -back-propagation through a recurrent neural network -which can make good on this promise. In particular, we discuss the universals of monotonicity, quantity, and conservativity and perform computational experiments of training such a network to learn to verify quantifiers. Our results are able to explain monotonicity and quantity quite well. We suggest that conservativity may have a different source than the other universals.

The vocabulary of human languages has been argued to support efficient communication by optimizing the trade-off between complexity and informativeness (Kemp & Regier 2012). The argument has been based on cross-linguistic analyses of vocabulary in semantic domains of content words such as kinship, color, and number terms. The present work extends this analysis to a category of function words: indefinite pronouns (e.g. someone, anyone, no-one, cf. Haspelmath 2001). We build on previous work to establish the meaning space and featural make-up for indefinite pronouns, and show that indefinite pronoun systems across languages optimize the complexity/informativeness trade-off. This demonstrates that pressures for efficient communication shape both content and function word categories, thus tying in with the conclusions of recent work on quantifiers by Steinert-Threlkeld (2019). Furthermore, we argue that the trade-off may explain some of the universal properties of indefinite pronouns, thus reducing the explanatory load for linguistic theories.

Although large-scale pretrained language models, such as BERT and RoBERTa, have achieved superhuman performance on indistribution test sets, their performance suffers on out-of-distribution test sets (e.g., on contrast sets). Building contrast sets often requires human-expert annotation, which is expensive and hard to create on a large scale. In this work, we propose a Linguistically-Informed Transformation (LIT) method to automatically generate contrast sets, which enables practitioners to explore linguistic phenomena of interests as well as compose different phenomena. Experimenting with our method on SNLI and MNLI shows that current pretrained language models, although being claimed to contain sufficient linguistic knowledge, struggle on our automatically generated contrast sets. Furthermore, we improve models' performance on the contrast sets by applying LIT to augment the training data, without affecting performance on the original data. 1

Expressivists about epistemic modals deny that 'Jane might be late' canonically serves to express the speaker's acceptance of a certain propositional content. Instead, they hold that it expresses a lack of acceptance (that Jane isn't late). Prominent expressivists embrace pragmatic expressivism: the doxastic property expressed by a declarative is not helpfully identified with (any part of) that sentence's compositional semantic value. Against this, we defend semantic expressivism about epistemic modals: the semantic value of a declarative from this domain is (partly) the property of doxastic attitudes it canonically serves to express. In support, we synthesize data from the critical literature on expressivism-largely reflecting interactions between modals and disjunctions-and present a semantic expressivism that readily predicts the data. This contrasts with salient competitors, including: pragmatic expressivism based on domain semantics or dynamic semantics; semantic expressivism à la Moss (Semant