To account for natural variability in cognitive processing, it is standard practice to optimize the parameters of a model to account for behavioral data. However, variability reflecting the information to which one has been exposed is usually ignored. Nevertheless, most language theories assign a large role to an individual’s experience with language. We present a new way to fit language-based behavioral data that combines simple learning and processing mechanisms using optimization of language materials. We demonstrate that benchmark fits on multiple linguistic tasks can be achieved using this method and will argue that one must account not only for the internal parameters of a model but also the external experience that people receive when theorizing about human behavior.

Distributional models of semantic cognition commonly makesimplifying assumptions, such as representing word co-occurrence structure by prototype-like high-dimensional se-mantic vectors, and limit how retrieval processes may con-tribute to the construction and use of semantic knowl-edge. More recently, the instance theory of semantics (ITS,Jamieson, Avery, Johns, & Jones, 2018) reconceived a dis-tributional model in terms of instance-based memory, allow-ing context-specific construction of semantic knowledge at thetime of retrieval. By simulation, we show that additional en-coding and retrieval operations, consistent with learning andmemory theory, can play a crucial role in flexibly controllingthe construction of general versus specific semantic knowl-edge. We argue this consolidation of processing principlesholds insight for distributional theories of semantic cognition.

Abstraction to a single prototypical representation is a core principle of Distributional Semantic Models (DSMs) that learn semantic representations for words by applying dimension reduction to statistical redundancies in language. While the learning mechanisms for semantic abstraction vary widely across the many DSMs in the literature, they are essentially all prototype models in that they create a single abstract representation for a word’s meaning. The prototype method stands in stark contrast to work in the field of categorization that has converged on the importance of instance models. In comparison to the prototype method, instance-based models assume only an episodic store and, rather than applying abstraction mechanisms at learning, argue that meaning emerges in the act of retrieval. We cash this idea out by presenting and evaluating an instance theory of distributional semantics, and by demonstrating that it can explain diverging patterns of homonymous words that classic “abstraction-at-learning” models simply cannot as a consequence of their architectural assumptions.

Recent research in the artificial grammar literature has found that a simple exemplar model of memory can account for a wide variety of artificial grammar results (Jamieson & Mewhort, 2009, 2010, 2011). This classic type of model has also been extended to account for natural language sentence processing effects (Johns & Jones, 2015). The current article extends this work to account for sentence production, and demonstrates that the structure of language itself provides sufficient power to generate syntactically correct sentences, even with no higher-level information about language provided to the model.