We examine the licensing of negative polarity items (NPIs) in large language models (LLMs) to enrich the picture of how models acquire NPIs as linguistic phenomena at the syntax-semantics interface. NPIs are a class of words which have a restricted distribution, appearing only in certain licensing contexts, prototypically negation. Unlike much of previous work which assumes NPIs and their licensing environments constitute unified classes, we consider NPI distribution in its full complexity: different NPIs are possible in different licensing environments. By studying this phenomenon across a broad range of models, we are able to explore which features of the model architecture, properties of the training data, and linguistic characteristics of the NPI phenomenon itself drive performance.

Language models are typically evaluated on their success at predicting the distribution of specific words in specific contexts. Yet linguistic knowledge also encodes relationships between contexts, allowing inferences between word distributions. We investigate the degree to which pre-trained transformer-based large language models (LLMs) represent such relationships, focusing on the domain of argument structure. We find that LLMs perform well in generalizing the distribution of a novel noun argument between related contexts that were seen during pre-training (e.g., the active object and passive subject of the verb spray), succeeding by making use of the semantically organized structure of the embedding space for word embeddings. However, LLMs fail at generalizations between related contexts that have not been observed during pre-training, but which instantiate more abstract, but well-attested structural generalizations (e.g., between the active object and passive subject of an arbitrary verb). Instead, in this case, LLMs show a bias to generalize based on linear order. This finding points to a limitation with current models and points to a reason for which their training is data-intensive.1

Due to the finite nature of any evidence used in learning, systematic generalization is crucially reliant on the presence of inductive bias (Mitchell, 1980). We examine inductive biases in different types of sequence-to-sequence neural network models, including CNNs, LSTMs (with and without attention), and transformers, inspired by Kharitonov and Chaabouni (2021). Crucially, however, we consider a wider range of possible inductive biases than their study did. Investigating preferences for hierarchical generalization compared to other types of generalization, we find that, contrary to their results, transformers display no preference for hierarchical generalization, but instead prefer a counting strategy. We also investigate biases toward different types of compositionality. By controlling for a confound in Kharitonov and Chaabouni (2021)’s test set, we find much less consistent generalization overall, and find that a large number of responses were among types other than the two types of generalization they had considered. Nevertheless, we observe consistent compositional generalization to held out combinations of primitives and functions on a SCAN task (Lake and Baroni, 2017) by models of all types, but only when primitives occur with other functions in the training set. The pattern of success indicates generalization in models of these types is highly sensitive to distributional properties of their training data.