On Semantic Cognition, Inductive Generalization, and Language Models

Our ability to understand language and perform reasoning crucially relies on a robust system of semantic cognition (G. L. Murphy, 2002; Rogers & McClelland, 2004; Rips et al., 2012; Lake & Murphy, 2021): processes that allow us to learn, update, and produce inferences about everyday concepts (e.g., cat, chair), properties (e.g., has fur, can be sat on), categories (e.g., mammals, furniture), and relations (e.g., is-a, taller-than). Meanwhile, recent progress in the field of natural language processing (NLP) has led to the development of language models (LMs): sophisticated neural networks that are trained to predict words in context (Devlin et al., 2019; Radford et al., 2019; Brown et al., 2020), and as a result build representations that encode the knowledge present in the statistics of their training environment. These models have achieved impressive levels of performance on a range of tasks that require sophisticated semantic knowledge (e.g. question answering and natural language inference), often even reaching human parity. To what extent do LMs capture the nuances of human conceptual knowledge and reasoning? Centering around this broad question, this dissertation uses core ideas in human semantic cognition as guiding principles and lays down the groundwork to establish effective evaluation and improvement of conceptual understanding in LMs. In particular, I build on prior work that focuses on characterizing what semantic knowledge is made available in the behavior and representations of LMs, and extend it by additionally proposing tests that focus on functional consequences of acquiring basic semantic knowledge.

I primarily focus on inductive generalization (Hayes & Heit, 2018)—the unique ability of humans to rely on acquired conceptual knowledge to project or generalize novel information—as a context within which we can analyze LMs’ encoding of conceptual knowledge. I do this, since the literature surrounding inductive generalization contains a variety of empirical regularities that map to specific conceptual abstractions and shed light on how humans store, organize and use conceptual knowledge. Before explicitly analyzing LMs for these empirical regularities, I test them on two other contexts, which also feature the role of inductive generalization. First I test the extent to which LMs demonstrate typicality effects—a robust finding in human categorization literature where certain members of a category are considered to be more central to the category than are others. Specifically, I test the behavior 19 different LMs on two contexts where typicality effects modulate human behavior: 1) verification of sentences expressing taxonomic category membership, and 2) projecting novel properties from individual category members to the entire category. In both tests, LMs achieved positive but modest correlations with human typicality ratings, suggesting that they can to a non-trivial extent capture subtle differences between category members. Next, I propose a new benchmark to test the robustness of LMs in attributing properties to everyday concepts, and in making inductive leaps to endow properties to novel concepts. On testing 31 different LMs for these capacities, I find that while they can correctly attribute properties to everyday concepts and even predict the properties of novel concepts in simple settings, they struggle to do so robustly. Combined with the analyses of typicality effects, these results suggest that the ability of LMs to demonstrate impressive conceptual knowledge and reasoning behavior can be explained by their sensitivities to shallow predictive cues. When these cues are carefully controlled for, LMs show critical failures in demonstrating robust conceptual understanding. Finally, I develop a framework that can allow us to characterize the extent to which the distributed representations learned by LMs can encode principles and abstractions that characterize inductive behavior of humans. This framework operationalizes inductive generalization as the behavior of an LM after its representations have been partially exposed (via gradient-based learning) to novel conceptual information. To simulate this behavior, the framework uses LMs that are endowed with human-elicited property knowledge, by training them to evaluate the truth of sentences attributing properties to concepts. I apply this framework to test four different LMs on 13 different inductive phenomena documented for humans (Osherson et al., 1990; Heit & Rubinstein, 1994). Results from these analyses suggest that building representations from word distributions can successfully allow the encoding of many abstract principles that can guide inductive behavior in the models—principles such as sensitivity to conceptual similarity, hierarchical organization of categories, reasoning about category coverage, and sample size. At the same time, the tested models also systematically failed at demonstrating certain phenomena, showcasing their inability to demonstrate pragmatic reasoning, preference to rely on shallow statistical cues, and lack of context sensitivity with respect to high-level intuitive theories.