To what extent can experience from language contribute to our conceptual knowledge? Computational explorations of this question have shed light on the ability of powerful neural language models (LMs)---informed solely through text input---to encode and elicit information about concepts and properties. To extend this line of research, we present a framework that uses neural-network language models (LMs) to perform property induction---a task in which humans generalize novel property knowledge (has sesamoid bones) from one or more concepts (robins) to others (sparrow, canary). Patterns of property induction observed in humans have shed considerable light on the nature and organization of human conceptual knowledge. Inspired by this insight, we use our framework to explore the property inductions of LMs, and find that they show an inductive preference to generalize novel properties on the basis of category membership, suggesting the presence of a taxonomic bias in their representations.

Building on research arguing for the possibility of conceptual and categorical knowledge acquisition through statistics contained in language, we evaluate predictive language models (LMs)---informed solely by textual input---on a prevalent phenomenon in cognitive science: typicality. Inspired by experiments that involve language processing and show robust typicality effects in humans, we propose two tests for LMs. Our first test targets whether typicality modulates LM probabilities in assigning taxonomic category memberships to items. The second test investigates sensitivities to typicality in LMs' probabilities when extending new information about items to their categories. Both tests show modest---but not completely absent---correspondence between LMs and humans, suggesting that text-based exposure alone is insufficient to acquire typicality knowledge.

The first language of a person has been shown to influence the processing of words when they learn a new language. Thishas been previously researched in behavioral studies, as well as by using lexical distributions or co-occurrence countsbetween word combinations to detect errors. In this paper we follow the findings of two recent studies and test theirhypotheses within the framework of two different word embedding models, based on their representation of the erroneoususage of concent words. Using an error-annotated corpus of essays written by learnings bellowing to 16 different firstlanguages, we compare incorrect words and their correct replacements as vectors in English and the learners first language.The results are consistent with previous findings that the first language has an influence on errors in the second language.The relationships between typologically similar languages differed between the models of embedding, suggesting anavenue for future explorations.

BERT is a language processing model trained for word prediction in context, which has shown impressive performancein natural language processing tasks. However, the principles underlying BERT’s use of linguistic cues present in contextare yet to be fully understood. In this work, we develop tests informed by the semantic priming paradigm to investigateBERTs handling of lexical relations to complete a cloze task (Taylor, 1953). We define priming to be an increase in BERTsexpectation for a target word (pilot), in a context (e.g., I want to be a ), when the context is prepended with a relatedword (airplane) as opposed to an unrelated one (table). We explore BERTs priming behavior under various predictiveconstraints placed on the blank, and find that BERT is sensitive to lexical priming effects only under minimal constraintfrom the input context. This pattern was found to be consistent across diverse lexical relations.