Modeling words in the mind

Constantine Lignos, University of Pennsylvania

Abstract

This dissertation uses computational modeling to address three related questions regarding the acquisition and processing of words: 1. How do infants learn to divide the speech stream into appropriately-sized words (word segmentation)? 2. How are these words represented and accessed in the adult mind? 3. How are generalizations about the regular and irregular characteristics of these words learned? Chapter 2 introduces a program for building computational models of cognition, discussing Marr's levels of analysis, the challenge of relating computer programs to cognitive processes, and the specific challenges introduced by modeling language and learning. I use the experiments reported by Saffran (2001) as a case study to demonstrate the value of using computational modeling even for simple learning experiments. Chapter 3 presents a model of infant word segmentation, developing an algorithmic-level model that mirrors the developmental trajectory of children as they learn to segment speech into words. The model proposed is computationally simple but is able to segment accurately when evaluated against child-directed speech corpora. Chapter 4 discusses the impasse that has formed regarding the processing of morphologically complex forms and presents the largest-scale study to date comparing dual-route and purely decompositional models of lexical access. I find that assuming obligatory decomposition of words into morphemes regardless of word frequency provides the best model of lexical decision data and allows for an algorithmic explanation of frequency effects using the rank hypothesis (Murray, 2004). Chapter 5 extends the findings regarding the processing of regular forms in Chapter 4 to the processing of irregular forms, finding support for a productivity-based model of irregular representation (Yang, 2005). I adopt this model and simulate a historical phonological change in the English language, postnasal stop deletion. By simulating learning across successive generations, I find that the productivity model replicates the historical trajectory and provides new insight as to what properties of English allowed the change to proceed and evidence that language learners are biased toward change, even when it results in added exceptionality. Chapter 6 concludes the dissertation, recapitulating the contributions of the previous chapters and suggesting avenues for future study

Subject Area

Linguistics|Cognitive psychology|Computer science

Recommended Citation

Lignos, Constantine, "Modeling words in the mind" (2013). Dissertations available from ProQuest. AAI3609197.
https://repository.upenn.edu/dissertations/AAI3609197

Share

COinS