Incremental planning and buffering in language production: Modeling large-scale data

Restricted (Penn State Only)
Cole, Jeremy
Graduate Program:
Information Sciences and Technology
Doctor of Philosophy
Document Type:
Date of Defense:
October 15, 2018
Committee Members:
  • David T Reitter, Dissertation Advisor
  • David T Reitter, Committee Chair
  • Clyde Lee Giles, Committee Member
  • Frank Edward Ritter, Committee Member
  • Michael Travis Putnam, Outside Member
  • computational linguistics
  • cognitive modeling
  • psycholinguistics
  • working memory
  • declarative memory
  • artificial intelligence
This dissertation sheds light on some of the representations used in human language production. Language production consists of several distinct stages; in particular, we focus on the stages of lexical retrieval and grammatical encoding. Lexical retrieval involves retrieving the words required to express the intended meaning; grammatical encoding concerns how to combine those words into an utterance. Here, we present studies relating declarative memory to working memory and buffering in fluent language production. In particular, both this dissertation and previous work provide evidence that some amount of language processing and planning is non-incremental. However, these subtle effects do not outweigh that grammatical encoding and buffering largely proceed incrementally. Indeed, there is no clear concrete strategy that can better explain the data than a strictly incremental one. This is likely because the type of non-incremental buffering we see is not something particularly explainable by a discrete strategy; instead, it is a complicated and nuanced strategy that depends on many external factors, including non-linguistic processing, such as semantic retrieval or reasoning. We build a series of models to investigate buffering in language production. First, a model of word adoption lays the groundwork for our computation of lexical retrieval from declarative memory. Next, we extend this to discuss a model of grammatical encoding that relies on Combinatory Categorial Grammar (CCG). The third study also builds on the model of declarative memory and extends it with a model of lexical retrieval that occurs across small time-scales. These three initial studies establish a modeling methodology that will inform the next two studies, which attempt to model lexical retrieval strategies from a discrete perspective and a differentiable one. In the end, our studies converge in arguing for mostly incremental planning and buffering of up to five words.